HW
H. Wang
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
65 records found
1
Private Cycle Enumeration in Federated Transaction Graphs
Collaborative Financial Crime Detection
The growth of digital financial systems has transformed both legitimate commerce and financial crime. Modern payment infrastructure enables fast cross-border transfers, global connectivity between institutions, and the automated processing of billions of transactions each year. While these developments have improved efficiency worldwide, they have also created new opportunities for criminals to conceal illicit activity through complex networks of accounts, institutions, and financial instruments. As a result, effective anti-money laundering enforcement increasingly depends on collaboration between organizations that each possess only a partial view of potentially suspicious activity. Such collaboration requires the exchange and joint analysis of highly sensitive financial data, creating tensions between security objectives, regulatory compliance, and individual privacy. The scale of modern financial systems further complicates analysis, motivating the use of distributed frameworks capable of processing massive datasets. Consequently, there is a growing need for privacy-preserving protocols that enable collaborative analysis at scale while revealing no more information than necessary.
One approach to financial crime detection models financial activity as a graph and searches for structural patterns that may indicate suspicious behavior. Among these patterns are cycles, where funds leave an account and return through a sequence of intermediary transactions. Detecting such subgraphs is challenging when transactions are distributed across multiple institutions, as no single party can observe them independently. Existing approaches either centralize data in a trusted location, raising privacy and trust concerns, or rely on decentralized computation among participants. In this thesis, we propose a decentralized, privacy-preserving cycle enumeration protocol for an arbitrary number of parties that finds all cycles of bounded length while minimizing information leakage. Our protocol combines homomorphic encryption, threshold cryptography, and secure multi-party computation, while building on offline graph processing techniques to improve scalability. We show that the protocol leaks less information than prior decentralized approaches, derive complexity bounds that exploit graph sparsity via out-degree dependence, and validate these results experimentally on graphs of varying sizes and types. We also discuss alternative design choices affecting security and performance. ...
One approach to financial crime detection models financial activity as a graph and searches for structural patterns that may indicate suspicious behavior. Among these patterns are cycles, where funds leave an account and return through a sequence of intermediary transactions. Detecting such subgraphs is challenging when transactions are distributed across multiple institutions, as no single party can observe them independently. Existing approaches either centralize data in a trusted location, raising privacy and trust concerns, or rely on decentralized computation among participants. In this thesis, we propose a decentralized, privacy-preserving cycle enumeration protocol for an arbitrary number of parties that finds all cycles of bounded length while minimizing information leakage. Our protocol combines homomorphic encryption, threshold cryptography, and secure multi-party computation, while building on offline graph processing techniques to improve scalability. We show that the protocol leaks less information than prior decentralized approaches, derive complexity bounds that exploit graph sparsity via out-degree dependence, and validate these results experimentally on graphs of varying sizes and types. We also discuss alternative design choices affecting security and performance. ...
The growth of digital financial systems has transformed both legitimate commerce and financial crime. Modern payment infrastructure enables fast cross-border transfers, global connectivity between institutions, and the automated processing of billions of transactions each year. While these developments have improved efficiency worldwide, they have also created new opportunities for criminals to conceal illicit activity through complex networks of accounts, institutions, and financial instruments. As a result, effective anti-money laundering enforcement increasingly depends on collaboration between organizations that each possess only a partial view of potentially suspicious activity. Such collaboration requires the exchange and joint analysis of highly sensitive financial data, creating tensions between security objectives, regulatory compliance, and individual privacy. The scale of modern financial systems further complicates analysis, motivating the use of distributed frameworks capable of processing massive datasets. Consequently, there is a growing need for privacy-preserving protocols that enable collaborative analysis at scale while revealing no more information than necessary.
One approach to financial crime detection models financial activity as a graph and searches for structural patterns that may indicate suspicious behavior. Among these patterns are cycles, where funds leave an account and return through a sequence of intermediary transactions. Detecting such subgraphs is challenging when transactions are distributed across multiple institutions, as no single party can observe them independently. Existing approaches either centralize data in a trusted location, raising privacy and trust concerns, or rely on decentralized computation among participants. In this thesis, we propose a decentralized, privacy-preserving cycle enumeration protocol for an arbitrary number of parties that finds all cycles of bounded length while minimizing information leakage. Our protocol combines homomorphic encryption, threshold cryptography, and secure multi-party computation, while building on offline graph processing techniques to improve scalability. We show that the protocol leaks less information than prior decentralized approaches, derive complexity bounds that exploit graph sparsity via out-degree dependence, and validate these results experimentally on graphs of varying sizes and types. We also discuss alternative design choices affecting security and performance.
One approach to financial crime detection models financial activity as a graph and searches for structural patterns that may indicate suspicious behavior. Among these patterns are cycles, where funds leave an account and return through a sequence of intermediary transactions. Detecting such subgraphs is challenging when transactions are distributed across multiple institutions, as no single party can observe them independently. Existing approaches either centralize data in a trusted location, raising privacy and trust concerns, or rely on decentralized computation among participants. In this thesis, we propose a decentralized, privacy-preserving cycle enumeration protocol for an arbitrary number of parties that finds all cycles of bounded length while minimizing information leakage. Our protocol combines homomorphic encryption, threshold cryptography, and secure multi-party computation, while building on offline graph processing techniques to improve scalability. We show that the protocol leaks less information than prior decentralized approaches, derive complexity bounds that exploit graph sparsity via out-degree dependence, and validate these results experimentally on graphs of varying sizes and types. We also discuss alternative design choices affecting security and performance.
Graphs are a powerful abstraction for reasoning about many real-world problems; however, it is often notoriously difficult to obtain the underlying graph representation directly. This is particularly relevant in applications such as epidemic response, where decisions such as vaccination depend on the connectivity of individuals, while the contact network itself may be only partially known. Diffusion network reconstruction addresses this problem by inferring the underlying graph from observations of a diffusion process occurring on it. Existing approaches typically aim to reconstruct a single most likely graph. Such a point estimate, however, does not capture uncertainty about alternative graph topologies that are also consistent with the available observations, which can be important for downstream decision making.
This thesis reformulates diffusion network reconstruction as a Bayesian inference problem by constructing a posterior distribution over possible graph topologies rather than a single point estimate. This allows downstream applications such as reinforcement learning (RL) to reason under uncertainty and make informed decisions when observations are limited or ambiguous. This thesis investigates the use of conditional generative flow networks (GFlowNets) to represent this posterior over graph topologies conditioned on observations from susceptible-infected (SI) epidemic simulations. To enable this approach, a likelihood-based reward function is introduced that factorizes the probability of an observed infection state and provides improved scalability over a naive empirical reward, achieving a true positive mass (TPM) of over 84% for graphs with up to 16 nodes while the baseline empirical reward function achieved a TPM of over 99% only up to 7 nodes. A tensor-optimized implementation of this reward function using PyTorch further reduces the computational cost of reward evaluation, achieving approximately 20 iterations per second compared with 2.1 iterations per second for a Julia implementation and 0.3 iterations per second for an implementation using NDlib under the evaluated configuration.
This thesis additionally investigates a custom linear transformer architecture designed to exploit the permutation structure of graph representations and improve the scalability of the conditional GFlowNet. The transformer achieves an improvement over a feed-forward architecture, reaching a 1-total variation distance (TVD) of 32% and a TPM of 49% on graphs with 25 nodes, compared with a 1-TVD of 12% and a TPM of 13% for a standard feed-forward architecture. However, this improvement comes at substantially increased computational cost, with the transformer requiring 86 seconds per forward pass compared with 16 seconds for the feed-forward model. Overall, the results demonstrate the feasibility of conditional GFlowNets for Bayesian diffusion network reconstruction on smaller graphs while the observed scalability limitations highlight reward evaluation, model scalability, and computational efficiency as key challenges for future work. ...
This thesis reformulates diffusion network reconstruction as a Bayesian inference problem by constructing a posterior distribution over possible graph topologies rather than a single point estimate. This allows downstream applications such as reinforcement learning (RL) to reason under uncertainty and make informed decisions when observations are limited or ambiguous. This thesis investigates the use of conditional generative flow networks (GFlowNets) to represent this posterior over graph topologies conditioned on observations from susceptible-infected (SI) epidemic simulations. To enable this approach, a likelihood-based reward function is introduced that factorizes the probability of an observed infection state and provides improved scalability over a naive empirical reward, achieving a true positive mass (TPM) of over 84% for graphs with up to 16 nodes while the baseline empirical reward function achieved a TPM of over 99% only up to 7 nodes. A tensor-optimized implementation of this reward function using PyTorch further reduces the computational cost of reward evaluation, achieving approximately 20 iterations per second compared with 2.1 iterations per second for a Julia implementation and 0.3 iterations per second for an implementation using NDlib under the evaluated configuration.
This thesis additionally investigates a custom linear transformer architecture designed to exploit the permutation structure of graph representations and improve the scalability of the conditional GFlowNet. The transformer achieves an improvement over a feed-forward architecture, reaching a 1-total variation distance (TVD) of 32% and a TPM of 49% on graphs with 25 nodes, compared with a 1-TVD of 12% and a TPM of 13% for a standard feed-forward architecture. However, this improvement comes at substantially increased computational cost, with the transformer requiring 86 seconds per forward pass compared with 16 seconds for the feed-forward model. Overall, the results demonstrate the feasibility of conditional GFlowNets for Bayesian diffusion network reconstruction on smaller graphs while the observed scalability limitations highlight reward evaluation, model scalability, and computational efficiency as key challenges for future work. ...
Graphs are a powerful abstraction for reasoning about many real-world problems; however, it is often notoriously difficult to obtain the underlying graph representation directly. This is particularly relevant in applications such as epidemic response, where decisions such as vaccination depend on the connectivity of individuals, while the contact network itself may be only partially known. Diffusion network reconstruction addresses this problem by inferring the underlying graph from observations of a diffusion process occurring on it. Existing approaches typically aim to reconstruct a single most likely graph. Such a point estimate, however, does not capture uncertainty about alternative graph topologies that are also consistent with the available observations, which can be important for downstream decision making.
This thesis reformulates diffusion network reconstruction as a Bayesian inference problem by constructing a posterior distribution over possible graph topologies rather than a single point estimate. This allows downstream applications such as reinforcement learning (RL) to reason under uncertainty and make informed decisions when observations are limited or ambiguous. This thesis investigates the use of conditional generative flow networks (GFlowNets) to represent this posterior over graph topologies conditioned on observations from susceptible-infected (SI) epidemic simulations. To enable this approach, a likelihood-based reward function is introduced that factorizes the probability of an observed infection state and provides improved scalability over a naive empirical reward, achieving a true positive mass (TPM) of over 84% for graphs with up to 16 nodes while the baseline empirical reward function achieved a TPM of over 99% only up to 7 nodes. A tensor-optimized implementation of this reward function using PyTorch further reduces the computational cost of reward evaluation, achieving approximately 20 iterations per second compared with 2.1 iterations per second for a Julia implementation and 0.3 iterations per second for an implementation using NDlib under the evaluated configuration.
This thesis additionally investigates a custom linear transformer architecture designed to exploit the permutation structure of graph representations and improve the scalability of the conditional GFlowNet. The transformer achieves an improvement over a feed-forward architecture, reaching a 1-total variation distance (TVD) of 32% and a TPM of 49% on graphs with 25 nodes, compared with a 1-TVD of 12% and a TPM of 13% for a standard feed-forward architecture. However, this improvement comes at substantially increased computational cost, with the transformer requiring 86 seconds per forward pass compared with 16 seconds for the feed-forward model. Overall, the results demonstrate the feasibility of conditional GFlowNets for Bayesian diffusion network reconstruction on smaller graphs while the observed scalability limitations highlight reward evaluation, model scalability, and computational efficiency as key challenges for future work.
This thesis reformulates diffusion network reconstruction as a Bayesian inference problem by constructing a posterior distribution over possible graph topologies rather than a single point estimate. This allows downstream applications such as reinforcement learning (RL) to reason under uncertainty and make informed decisions when observations are limited or ambiguous. This thesis investigates the use of conditional generative flow networks (GFlowNets) to represent this posterior over graph topologies conditioned on observations from susceptible-infected (SI) epidemic simulations. To enable this approach, a likelihood-based reward function is introduced that factorizes the probability of an observed infection state and provides improved scalability over a naive empirical reward, achieving a true positive mass (TPM) of over 84% for graphs with up to 16 nodes while the baseline empirical reward function achieved a TPM of over 99% only up to 7 nodes. A tensor-optimized implementation of this reward function using PyTorch further reduces the computational cost of reward evaluation, achieving approximately 20 iterations per second compared with 2.1 iterations per second for a Julia implementation and 0.3 iterations per second for an implementation using NDlib under the evaluated configuration.
This thesis additionally investigates a custom linear transformer architecture designed to exploit the permutation structure of graph representations and improve the scalability of the conditional GFlowNet. The transformer achieves an improvement over a feed-forward architecture, reaching a 1-total variation distance (TVD) of 32% and a TPM of 49% on graphs with 25 nodes, compared with a 1-TVD of 12% and a TPM of 13% for a standard feed-forward architecture. However, this improvement comes at substantially increased computational cost, with the transformer requiring 86 seconds per forward pass compared with 16 seconds for the feed-forward model. Overall, the results demonstrate the feasibility of conditional GFlowNets for Bayesian diffusion network reconstruction on smaller graphs while the observed scalability limitations highlight reward evaluation, model scalability, and computational efficiency as key challenges for future work.
The rapid growth of mobile data traffic and the evolution towards 5G-Advanced and 6G networks have significantly increased the operational complexity of mobile networks, making fault localization a critical challenge for mobile network operators. Traditional fault management approaches rely on reactive, threshold-based alarms operating within individual network domains. This is increasingly ineffective for detecting and localizing faults in complex, multi-domain environments.
This thesis proposes a cross-domain fault localization framework that combines unsupervised anomaly detection with offline reinforcement learning, where the RL agent learns a policy for identifying the most likely fault origin based on observed anomaly patterns across domains. The proposed framework analyzes time-series Key Performance Indicators (KPIs) collected from the Radio Access Network (RAN), Core Network and end-to-end (E2E) domains to detect anomalous behaviour without requiring labeled data. Subsequently, the detected anomalies and cross-domain results are used by an offline reinforcement learning agent to track the most probable origin of faults across network domains.
The framework is evaluated using real-world KPI data collected over a 1 month period, consisting of 13 KPIs across the 3 domains. Unsupervised anomaly detection is applied to identify deviations from normal network behavior, while fault localization is performed using reinforcement learning based on the observed anomaly patterns. The results demonstrate that the proposed approach can identify anomalies and provide fault localization despite the lack of explicit fault labels.
This thesis highlights key challenges such as traffic dependent KPI behaviour, noise during low traffic periods and threshold selection. While the results indicate that combining unsupervised anomaly detection with reinforcement learning is a promising direction for predictive fault localization, further refinement is required to improve robustness, precision and operational consistency. Future work should focus on online deployment, calibration and improved learning strategies to support deployment in real world mobile network environments. ...
This thesis proposes a cross-domain fault localization framework that combines unsupervised anomaly detection with offline reinforcement learning, where the RL agent learns a policy for identifying the most likely fault origin based on observed anomaly patterns across domains. The proposed framework analyzes time-series Key Performance Indicators (KPIs) collected from the Radio Access Network (RAN), Core Network and end-to-end (E2E) domains to detect anomalous behaviour without requiring labeled data. Subsequently, the detected anomalies and cross-domain results are used by an offline reinforcement learning agent to track the most probable origin of faults across network domains.
The framework is evaluated using real-world KPI data collected over a 1 month period, consisting of 13 KPIs across the 3 domains. Unsupervised anomaly detection is applied to identify deviations from normal network behavior, while fault localization is performed using reinforcement learning based on the observed anomaly patterns. The results demonstrate that the proposed approach can identify anomalies and provide fault localization despite the lack of explicit fault labels.
This thesis highlights key challenges such as traffic dependent KPI behaviour, noise during low traffic periods and threshold selection. While the results indicate that combining unsupervised anomaly detection with reinforcement learning is a promising direction for predictive fault localization, further refinement is required to improve robustness, precision and operational consistency. Future work should focus on online deployment, calibration and improved learning strategies to support deployment in real world mobile network environments. ...
The rapid growth of mobile data traffic and the evolution towards 5G-Advanced and 6G networks have significantly increased the operational complexity of mobile networks, making fault localization a critical challenge for mobile network operators. Traditional fault management approaches rely on reactive, threshold-based alarms operating within individual network domains. This is increasingly ineffective for detecting and localizing faults in complex, multi-domain environments.
This thesis proposes a cross-domain fault localization framework that combines unsupervised anomaly detection with offline reinforcement learning, where the RL agent learns a policy for identifying the most likely fault origin based on observed anomaly patterns across domains. The proposed framework analyzes time-series Key Performance Indicators (KPIs) collected from the Radio Access Network (RAN), Core Network and end-to-end (E2E) domains to detect anomalous behaviour without requiring labeled data. Subsequently, the detected anomalies and cross-domain results are used by an offline reinforcement learning agent to track the most probable origin of faults across network domains.
The framework is evaluated using real-world KPI data collected over a 1 month period, consisting of 13 KPIs across the 3 domains. Unsupervised anomaly detection is applied to identify deviations from normal network behavior, while fault localization is performed using reinforcement learning based on the observed anomaly patterns. The results demonstrate that the proposed approach can identify anomalies and provide fault localization despite the lack of explicit fault labels.
This thesis highlights key challenges such as traffic dependent KPI behaviour, noise during low traffic periods and threshold selection. While the results indicate that combining unsupervised anomaly detection with reinforcement learning is a promising direction for predictive fault localization, further refinement is required to improve robustness, precision and operational consistency. Future work should focus on online deployment, calibration and improved learning strategies to support deployment in real world mobile network environments.
This thesis proposes a cross-domain fault localization framework that combines unsupervised anomaly detection with offline reinforcement learning, where the RL agent learns a policy for identifying the most likely fault origin based on observed anomaly patterns across domains. The proposed framework analyzes time-series Key Performance Indicators (KPIs) collected from the Radio Access Network (RAN), Core Network and end-to-end (E2E) domains to detect anomalous behaviour without requiring labeled data. Subsequently, the detected anomalies and cross-domain results are used by an offline reinforcement learning agent to track the most probable origin of faults across network domains.
The framework is evaluated using real-world KPI data collected over a 1 month period, consisting of 13 KPIs across the 3 domains. Unsupervised anomaly detection is applied to identify deviations from normal network behavior, while fault localization is performed using reinforcement learning based on the observed anomaly patterns. The results demonstrate that the proposed approach can identify anomalies and provide fault localization despite the lack of explicit fault labels.
This thesis highlights key challenges such as traffic dependent KPI behaviour, noise during low traffic periods and threshold selection. While the results indicate that combining unsupervised anomaly detection with reinforcement learning is a promising direction for predictive fault localization, further refinement is required to improve robustness, precision and operational consistency. Future work should focus on online deployment, calibration and improved learning strategies to support deployment in real world mobile network environments.
Blockchain technology has significantly impacted digital transactions and data management by providing a decentralized, transparent, and immutable ledger. Private blockchains, unlike public blockchains, are restricted to a pre-selected group of participants, making them more suitable for controlled environments such as enterprises, governments, or academic institutions. Hyperledger Fabric (HLF) is a widely used framework for private blockchain technology, designed for enterprise use.
With quantum computers on the rise, commonly used cryptographic algorithms are increasingly at risk of becoming obsolete. Blockchain networks rely extensively on these primitives, making them particularly vulnerable to advances in quantum computing. To counter this vulnerability, post-quantum algorithms have gained popularity within the cryptographic community.
This thesis focuses on securing private blockchains built on HLF against potential quantum adversaries using post-quantum cryptographic primitives. We implement ML-DSA, Vesper, and TDUE as smart contracts for digital signing, zero-knowledge proofs, and updatable encryption, respectively, and report on their performance. Furthermore, we build on top of Fabric Private Chaincode to maintain the confidentiality of the contract application state. While overall performance is not yet competitive with classical cryptographic primitives, our findings indicate that post-quantum primitives have promising potential for use in private blockchains. ...
With quantum computers on the rise, commonly used cryptographic algorithms are increasingly at risk of becoming obsolete. Blockchain networks rely extensively on these primitives, making them particularly vulnerable to advances in quantum computing. To counter this vulnerability, post-quantum algorithms have gained popularity within the cryptographic community.
This thesis focuses on securing private blockchains built on HLF against potential quantum adversaries using post-quantum cryptographic primitives. We implement ML-DSA, Vesper, and TDUE as smart contracts for digital signing, zero-knowledge proofs, and updatable encryption, respectively, and report on their performance. Furthermore, we build on top of Fabric Private Chaincode to maintain the confidentiality of the contract application state. While overall performance is not yet competitive with classical cryptographic primitives, our findings indicate that post-quantum primitives have promising potential for use in private blockchains. ...
Blockchain technology has significantly impacted digital transactions and data management by providing a decentralized, transparent, and immutable ledger. Private blockchains, unlike public blockchains, are restricted to a pre-selected group of participants, making them more suitable for controlled environments such as enterprises, governments, or academic institutions. Hyperledger Fabric (HLF) is a widely used framework for private blockchain technology, designed for enterprise use.
With quantum computers on the rise, commonly used cryptographic algorithms are increasingly at risk of becoming obsolete. Blockchain networks rely extensively on these primitives, making them particularly vulnerable to advances in quantum computing. To counter this vulnerability, post-quantum algorithms have gained popularity within the cryptographic community.
This thesis focuses on securing private blockchains built on HLF against potential quantum adversaries using post-quantum cryptographic primitives. We implement ML-DSA, Vesper, and TDUE as smart contracts for digital signing, zero-knowledge proofs, and updatable encryption, respectively, and report on their performance. Furthermore, we build on top of Fabric Private Chaincode to maintain the confidentiality of the contract application state. While overall performance is not yet competitive with classical cryptographic primitives, our findings indicate that post-quantum primitives have promising potential for use in private blockchains.
With quantum computers on the rise, commonly used cryptographic algorithms are increasingly at risk of becoming obsolete. Blockchain networks rely extensively on these primitives, making them particularly vulnerable to advances in quantum computing. To counter this vulnerability, post-quantum algorithms have gained popularity within the cryptographic community.
This thesis focuses on securing private blockchains built on HLF against potential quantum adversaries using post-quantum cryptographic primitives. We implement ML-DSA, Vesper, and TDUE as smart contracts for digital signing, zero-knowledge proofs, and updatable encryption, respectively, and report on their performance. Furthermore, we build on top of Fabric Private Chaincode to maintain the confidentiality of the contract application state. While overall performance is not yet competitive with classical cryptographic primitives, our findings indicate that post-quantum primitives have promising potential for use in private blockchains.
Appearance-based 3D gaze estimation must accommodate two conflicting needs: fine ocular detail and global facial context. Vanilla Vision Transformers (ViTs) struggle with both needs due to their fixed 16 × 16 patch grid that (i) fragments critical features like the eyes into multiple patches, and (ii) floods the self-attention mechanism with redundant information from the forehead, cheeks and background. We introduce FocusViT, a lightweight and end-to-end differentiable framework that enhances ViTs by first using a Patch Translation Module, based on SpatialTransformer Networks, to dynamically translate patches to center on content, and then employing a Perturbed Top-K operator to select only the most informative tokens for processing.
Our experiments show that combining translation and selection reduces the mean angular error (MAE) of a ViTS baseline on ETH-XGaze from 4.98◦ to 4.61◦ while using 75% fewer tokens. Furthermore, by leveraging this token reduction to enable a finer-grained, lossless 8x8 patch grid, we address a key information bottleneck in the ViT-S architecture, achieving a final MAE of 4.42◦. The framework also demonstrates consistent improvements on the MPIIFaceGaze dataset, reducing the baseline error from 5.72◦ to 5.36◦. Extensive ablation studies confirm our central finding: patch translation and token selection are complementary mechanisms that work in synergy to improve model performance. ...
Our experiments show that combining translation and selection reduces the mean angular error (MAE) of a ViTS baseline on ETH-XGaze from 4.98◦ to 4.61◦ while using 75% fewer tokens. Furthermore, by leveraging this token reduction to enable a finer-grained, lossless 8x8 patch grid, we address a key information bottleneck in the ViT-S architecture, achieving a final MAE of 4.42◦. The framework also demonstrates consistent improvements on the MPIIFaceGaze dataset, reducing the baseline error from 5.72◦ to 5.36◦. Extensive ablation studies confirm our central finding: patch translation and token selection are complementary mechanisms that work in synergy to improve model performance. ...
Appearance-based 3D gaze estimation must accommodate two conflicting needs: fine ocular detail and global facial context. Vanilla Vision Transformers (ViTs) struggle with both needs due to their fixed 16 × 16 patch grid that (i) fragments critical features like the eyes into multiple patches, and (ii) floods the self-attention mechanism with redundant information from the forehead, cheeks and background. We introduce FocusViT, a lightweight and end-to-end differentiable framework that enhances ViTs by first using a Patch Translation Module, based on SpatialTransformer Networks, to dynamically translate patches to center on content, and then employing a Perturbed Top-K operator to select only the most informative tokens for processing.
Our experiments show that combining translation and selection reduces the mean angular error (MAE) of a ViTS baseline on ETH-XGaze from 4.98◦ to 4.61◦ while using 75% fewer tokens. Furthermore, by leveraging this token reduction to enable a finer-grained, lossless 8x8 patch grid, we address a key information bottleneck in the ViT-S architecture, achieving a final MAE of 4.42◦. The framework also demonstrates consistent improvements on the MPIIFaceGaze dataset, reducing the baseline error from 5.72◦ to 5.36◦. Extensive ablation studies confirm our central finding: patch translation and token selection are complementary mechanisms that work in synergy to improve model performance.
Our experiments show that combining translation and selection reduces the mean angular error (MAE) of a ViTS baseline on ETH-XGaze from 4.98◦ to 4.61◦ while using 75% fewer tokens. Furthermore, by leveraging this token reduction to enable a finer-grained, lossless 8x8 patch grid, we address a key information bottleneck in the ViT-S architecture, achieving a final MAE of 4.42◦. The framework also demonstrates consistent improvements on the MPIIFaceGaze dataset, reducing the baseline error from 5.72◦ to 5.36◦. Extensive ablation studies confirm our central finding: patch translation and token selection are complementary mechanisms that work in synergy to improve model performance.
Contract bridge is a challenging card game that combines strategic bidding, teamwork, and probabilistic outcomes. This thesis develops a rigorous mathematical framework for analyzing duplicate bridge tournaments, with a focus on fair and accurate player ranking. We introduce a threshold ranking model that extends the classic Bradley–Terry model by incorporating a tie parameter, allowing us to model the non-negligible probability of tied scores in bridge results. We derive theoretical results for this model, notably, we prove that the total match-points and number of tied boards are sufficient statistics for the estimation of players’ skill levels and the tie parameter. Using both a reference dataset from literature and the results of the 2024 China National Bridge Championship, we demonstrate that our model can reliably estimate skill parameters: the inferred rankings align with actual tournament outcomes and distinguish the strongest and weakest pairs. To complement the parametric model, we employ a fuzzy logic approach to categorize board-by-board performance into qualitative grades (“Excellent”, “Good”, etc.) and compute aggregate fuzzy performance scores. This provides an additional lens to evaluate each pair’s consistency and tendency for extreme results. The combined analysis yields a richer understanding of performance: while the statistical model emphasizes overall consistency and head-to-head advantages, the fuzzy analysis highlights variability and exceptional highs or lows. Our results show that both methods concur on the identification of top performers, reinforcing confidence in the findings, and together they reveal nuanced differences among closely ranked competitors. These insights further confirm the fairness of the duplicate format (skill prevails in the long run) while quantifying the impact of chance and strategic diversity on interim rankings.
In summary, the contributions of this thesis are: (1) a new tie-aware mathematical model for ranking players in duplicate bridge, supported by theoretical guarantees and validated by empirical data; (2) the integration of fuzzy logic metrics into bridge performance evaluation, offering practical measures of consistency; and (3) a comprehensive case study on championship data illustrating how our methods can be applied to draw meaningful conclusions about player abilities and game dynamics. This work not only advances the methodology for analyzing bridge tournaments but also provides tools and perspectives that can inform the design of fair competition systems and the training of competitive bridge players.
...
In summary, the contributions of this thesis are: (1) a new tie-aware mathematical model for ranking players in duplicate bridge, supported by theoretical guarantees and validated by empirical data; (2) the integration of fuzzy logic metrics into bridge performance evaluation, offering practical measures of consistency; and (3) a comprehensive case study on championship data illustrating how our methods can be applied to draw meaningful conclusions about player abilities and game dynamics. This work not only advances the methodology for analyzing bridge tournaments but also provides tools and perspectives that can inform the design of fair competition systems and the training of competitive bridge players.
...
Contract bridge is a challenging card game that combines strategic bidding, teamwork, and probabilistic outcomes. This thesis develops a rigorous mathematical framework for analyzing duplicate bridge tournaments, with a focus on fair and accurate player ranking. We introduce a threshold ranking model that extends the classic Bradley–Terry model by incorporating a tie parameter, allowing us to model the non-negligible probability of tied scores in bridge results. We derive theoretical results for this model, notably, we prove that the total match-points and number of tied boards are sufficient statistics for the estimation of players’ skill levels and the tie parameter. Using both a reference dataset from literature and the results of the 2024 China National Bridge Championship, we demonstrate that our model can reliably estimate skill parameters: the inferred rankings align with actual tournament outcomes and distinguish the strongest and weakest pairs. To complement the parametric model, we employ a fuzzy logic approach to categorize board-by-board performance into qualitative grades (“Excellent”, “Good”, etc.) and compute aggregate fuzzy performance scores. This provides an additional lens to evaluate each pair’s consistency and tendency for extreme results. The combined analysis yields a richer understanding of performance: while the statistical model emphasizes overall consistency and head-to-head advantages, the fuzzy analysis highlights variability and exceptional highs or lows. Our results show that both methods concur on the identification of top performers, reinforcing confidence in the findings, and together they reveal nuanced differences among closely ranked competitors. These insights further confirm the fairness of the duplicate format (skill prevails in the long run) while quantifying the impact of chance and strategic diversity on interim rankings.
In summary, the contributions of this thesis are: (1) a new tie-aware mathematical model for ranking players in duplicate bridge, supported by theoretical guarantees and validated by empirical data; (2) the integration of fuzzy logic metrics into bridge performance evaluation, offering practical measures of consistency; and (3) a comprehensive case study on championship data illustrating how our methods can be applied to draw meaningful conclusions about player abilities and game dynamics. This work not only advances the methodology for analyzing bridge tournaments but also provides tools and perspectives that can inform the design of fair competition systems and the training of competitive bridge players.
In summary, the contributions of this thesis are: (1) a new tie-aware mathematical model for ranking players in duplicate bridge, supported by theoretical guarantees and validated by empirical data; (2) the integration of fuzzy logic metrics into bridge performance evaluation, offering practical measures of consistency; and (3) a comprehensive case study on championship data illustrating how our methods can be applied to draw meaningful conclusions about player abilities and game dynamics. This work not only advances the methodology for analyzing bridge tournaments but also provides tools and perspectives that can inform the design of fair competition systems and the training of competitive bridge players.
Audio-to-motion generation is an important task with applications in virtual avatar creation for XR systems and intelligent robot control in daily life scenarios.
Most current motion generation methods depend on a single encoder-decoder architecture to simultaneously model all body parts, constraining their capacity to capture the diverse and complex motions exhibited by humans.
In this paper, we propose a novel method, SpeechCAT, that employs three separate encoder-decoder modules to individually model the motions of the face, body, and hands. To capture the relationships and synchronization among these body parts, we introduce a cross-attention mechanism to effectively learn their correlations.
SpeechCAT ensures sufficient capacity to model the unique characteristics of each body part while preserving the coherence between them.
Our experimental results demonstrate the superiority of SpeechCAT over baseline methods, highlighting its effectiveness in generating diverse, realistic, and synchronized motions with face, body, and hand parts. ...
Most current motion generation methods depend on a single encoder-decoder architecture to simultaneously model all body parts, constraining their capacity to capture the diverse and complex motions exhibited by humans.
In this paper, we propose a novel method, SpeechCAT, that employs three separate encoder-decoder modules to individually model the motions of the face, body, and hands. To capture the relationships and synchronization among these body parts, we introduce a cross-attention mechanism to effectively learn their correlations.
SpeechCAT ensures sufficient capacity to model the unique characteristics of each body part while preserving the coherence between them.
Our experimental results demonstrate the superiority of SpeechCAT over baseline methods, highlighting its effectiveness in generating diverse, realistic, and synchronized motions with face, body, and hand parts. ...
Audio-to-motion generation is an important task with applications in virtual avatar creation for XR systems and intelligent robot control in daily life scenarios.
Most current motion generation methods depend on a single encoder-decoder architecture to simultaneously model all body parts, constraining their capacity to capture the diverse and complex motions exhibited by humans.
In this paper, we propose a novel method, SpeechCAT, that employs three separate encoder-decoder modules to individually model the motions of the face, body, and hands. To capture the relationships and synchronization among these body parts, we introduce a cross-attention mechanism to effectively learn their correlations.
SpeechCAT ensures sufficient capacity to model the unique characteristics of each body part while preserving the coherence between them.
Our experimental results demonstrate the superiority of SpeechCAT over baseline methods, highlighting its effectiveness in generating diverse, realistic, and synchronized motions with face, body, and hand parts.
Most current motion generation methods depend on a single encoder-decoder architecture to simultaneously model all body parts, constraining their capacity to capture the diverse and complex motions exhibited by humans.
In this paper, we propose a novel method, SpeechCAT, that employs three separate encoder-decoder modules to individually model the motions of the face, body, and hands. To capture the relationships and synchronization among these body parts, we introduce a cross-attention mechanism to effectively learn their correlations.
SpeechCAT ensures sufficient capacity to model the unique characteristics of each body part while preserving the coherence between them.
Our experimental results demonstrate the superiority of SpeechCAT over baseline methods, highlighting its effectiveness in generating diverse, realistic, and synchronized motions with face, body, and hand parts.
Many fields rely on scarce and sensitive time series data, such as patient health records. Privacy regulations often make sharing such data challenging, slowing research progress. Synthetic time series offer a potential solution by replicating statistical characteristics of real data without revealing private information. Yet, they introduce new risks, as synthetic data may be mistaken for real. Watermarking can mitigate this by embedding a machine-detectable signal that preserves data quality. For such methods to be effective, watermarks must be robust to removal attempts. Existing research lacks direct comparisons of generative models for time series synthesis and watermarking. Furthermore, they only evaluate watermark robustness against time-domain attacks. Attacks in other domains, such as the frequency domain, remain unexplored. In order to address these gaps, this thesis investigates three key questions. First, which generative models are best suited for time series synthesis. Second, whether latent diffusion models (LDMs) can support watermarking. Lastly, how robust existing diffusion watermarks are against adversarial attacks.
A comparative study between GPT-based models and diffusion models showed that diffusion models produce synthetic data of higher quality. LDMs were then evaluated as a potential alternative. Their reliance on a variational autoencoder led to low quality outputs. Hence, standard diffusion models were elected as the superior watermarking candidate. Finally, we introduced an extended set of time-, frequency-, and time-frequency domain attacks to asses watermark robustness. TimeWak emerged as the most robust watermark. However, our extended attack suite revealed new vulnerabilities in all watermarks, highlighting the importance of comprehensive robustness evaluations. ...
A comparative study between GPT-based models and diffusion models showed that diffusion models produce synthetic data of higher quality. LDMs were then evaluated as a potential alternative. Their reliance on a variational autoencoder led to low quality outputs. Hence, standard diffusion models were elected as the superior watermarking candidate. Finally, we introduced an extended set of time-, frequency-, and time-frequency domain attacks to asses watermark robustness. TimeWak emerged as the most robust watermark. However, our extended attack suite revealed new vulnerabilities in all watermarks, highlighting the importance of comprehensive robustness evaluations. ...
Many fields rely on scarce and sensitive time series data, such as patient health records. Privacy regulations often make sharing such data challenging, slowing research progress. Synthetic time series offer a potential solution by replicating statistical characteristics of real data without revealing private information. Yet, they introduce new risks, as synthetic data may be mistaken for real. Watermarking can mitigate this by embedding a machine-detectable signal that preserves data quality. For such methods to be effective, watermarks must be robust to removal attempts. Existing research lacks direct comparisons of generative models for time series synthesis and watermarking. Furthermore, they only evaluate watermark robustness against time-domain attacks. Attacks in other domains, such as the frequency domain, remain unexplored. In order to address these gaps, this thesis investigates three key questions. First, which generative models are best suited for time series synthesis. Second, whether latent diffusion models (LDMs) can support watermarking. Lastly, how robust existing diffusion watermarks are against adversarial attacks.
A comparative study between GPT-based models and diffusion models showed that diffusion models produce synthetic data of higher quality. LDMs were then evaluated as a potential alternative. Their reliance on a variational autoencoder led to low quality outputs. Hence, standard diffusion models were elected as the superior watermarking candidate. Finally, we introduced an extended set of time-, frequency-, and time-frequency domain attacks to asses watermark robustness. TimeWak emerged as the most robust watermark. However, our extended attack suite revealed new vulnerabilities in all watermarks, highlighting the importance of comprehensive robustness evaluations.
A comparative study between GPT-based models and diffusion models showed that diffusion models produce synthetic data of higher quality. LDMs were then evaluated as a potential alternative. Their reliance on a variational autoencoder led to low quality outputs. Hence, standard diffusion models were elected as the superior watermarking candidate. Finally, we introduced an extended set of time-, frequency-, and time-frequency domain attacks to asses watermark robustness. TimeWak emerged as the most robust watermark. However, our extended attack suite revealed new vulnerabilities in all watermarks, highlighting the importance of comprehensive robustness evaluations.
Understanding Users’ Contextual Factors and Personal Values for Watching YouTube Videos
A Crowdsourcing Approach with Personal Reflection Integration
User feedback plays a significant role in helping recommendation systems to make personalized and accurate predictions. Despite the fact that many methods of collecting user feedback have been proposed, little research exists that addresses both the breadth and depth of data collected. In this study, we incorporate personal reflection into traditional crowdsourcing tasks and investigate how it facilitates people's reflection on their usage of YouTube. We present a novel crowdsourcing approach with personal reflection integration based on several design principles, which allows participants to reflect on contextual factors and personal values of using YouTube through guided context recall and evaluations, therefore gathering deeper insights into people's preferences and behaviors on a large scale. We conducted a user study involving 20 participants and explored the insights generated from their textual answers and the role of design principles in the reflection process. This approach successfully enables multiple participants to conduct the study simultaneously, thereby reflecting on their watching behaviors and preferences. Based on the quantitative and qualitative analysis of the findings, we sum up the implications of this approach to provide guidance for the YouTube recommendation system and point to the directions for the design of similar studies in the future.
...
User feedback plays a significant role in helping recommendation systems to make personalized and accurate predictions. Despite the fact that many methods of collecting user feedback have been proposed, little research exists that addresses both the breadth and depth of data collected. In this study, we incorporate personal reflection into traditional crowdsourcing tasks and investigate how it facilitates people's reflection on their usage of YouTube. We present a novel crowdsourcing approach with personal reflection integration based on several design principles, which allows participants to reflect on contextual factors and personal values of using YouTube through guided context recall and evaluations, therefore gathering deeper insights into people's preferences and behaviors on a large scale. We conducted a user study involving 20 participants and explored the insights generated from their textual answers and the role of design principles in the reflection process. This approach successfully enables multiple participants to conduct the study simultaneously, thereby reflecting on their watching behaviors and preferences. Based on the quantitative and qualitative analysis of the findings, we sum up the implications of this approach to provide guidance for the YouTube recommendation system and point to the directions for the design of similar studies in the future.
This thesis investigates the potential of solving large-scale linear systems of equations (LSEs) in the Tensor Train (TT) format. First, we study when this approach can outperform traditional solvers like Conjugate Gradient (CG) and Gaussian Elimination (GE). In turn, we examine the time complexity of the TT-solve technique when applied to ill-conditioned, sparse, and symmetric positive definite (SPD) matrices. This enables us to assess the scenarios where TT-solve may offer significant advantages.
During this exploration, we have also identified and proposed solutions for a handful of research gaps that could further optimize the TT-solve process. These include challenges related to poorly factorizing matrix sizes, as well as the need for a deeper understanding of the trade-offs between TT-matrix (TTM) rank and maximal mode size during the decomposition process. Additionally, the study examined the potential for altering the coefficient matrix through techniques like variable reordering and partial Gaussian elimination (PG) to achieve lower TTM-ranks without affecting the solution of the LSE. ...
During this exploration, we have also identified and proposed solutions for a handful of research gaps that could further optimize the TT-solve process. These include challenges related to poorly factorizing matrix sizes, as well as the need for a deeper understanding of the trade-offs between TT-matrix (TTM) rank and maximal mode size during the decomposition process. Additionally, the study examined the potential for altering the coefficient matrix through techniques like variable reordering and partial Gaussian elimination (PG) to achieve lower TTM-ranks without affecting the solution of the LSE. ...
This thesis investigates the potential of solving large-scale linear systems of equations (LSEs) in the Tensor Train (TT) format. First, we study when this approach can outperform traditional solvers like Conjugate Gradient (CG) and Gaussian Elimination (GE). In turn, we examine the time complexity of the TT-solve technique when applied to ill-conditioned, sparse, and symmetric positive definite (SPD) matrices. This enables us to assess the scenarios where TT-solve may offer significant advantages.
During this exploration, we have also identified and proposed solutions for a handful of research gaps that could further optimize the TT-solve process. These include challenges related to poorly factorizing matrix sizes, as well as the need for a deeper understanding of the trade-offs between TT-matrix (TTM) rank and maximal mode size during the decomposition process. Additionally, the study examined the potential for altering the coefficient matrix through techniques like variable reordering and partial Gaussian elimination (PG) to achieve lower TTM-ranks without affecting the solution of the LSE.
During this exploration, we have also identified and proposed solutions for a handful of research gaps that could further optimize the TT-solve process. These include challenges related to poorly factorizing matrix sizes, as well as the need for a deeper understanding of the trade-offs between TT-matrix (TTM) rank and maximal mode size during the decomposition process. Additionally, the study examined the potential for altering the coefficient matrix through techniques like variable reordering and partial Gaussian elimination (PG) to achieve lower TTM-ranks without affecting the solution of the LSE.
Cluster analysis in high dimensional data is a difficult but desirable task. Many existing methods fail to cluster high dimensional data due to what is known as the curse of dimensionality. Therefore, sophisticated clustering methods are in wide development. Along these lines, spectral modularity maximization emerged from the theory of random matrices and graph modularity. The method is based on a filtering of the spectral decomposition of similarity matrices. Despite the recent success of this method, we uncover a fundamental challenge of spectral modularity: the spectral modularity breaks down as the number of groups in a data set grows. To mitigate this challenge, we propose two solutions: one solution based on a regularization and one solution based on a normalization. We perform a thorough empirical analysis of the clustering performance of the solutions and find that, not only do our methods resolve the breakdown of spectral modularity, but they also outperform existing clustering methods in
a variety of settings. ...
a variety of settings. ...
Cluster analysis in high dimensional data is a difficult but desirable task. Many existing methods fail to cluster high dimensional data due to what is known as the curse of dimensionality. Therefore, sophisticated clustering methods are in wide development. Along these lines, spectral modularity maximization emerged from the theory of random matrices and graph modularity. The method is based on a filtering of the spectral decomposition of similarity matrices. Despite the recent success of this method, we uncover a fundamental challenge of spectral modularity: the spectral modularity breaks down as the number of groups in a data set grows. To mitigate this challenge, we propose two solutions: one solution based on a regularization and one solution based on a normalization. We perform a thorough empirical analysis of the clustering performance of the solutions and find that, not only do our methods resolve the breakdown of spectral modularity, but they also outperform existing clustering methods in
a variety of settings.
a variety of settings.
Examining the Efficacy of Persuasive eHealth Applications in Facilitating Smoking Cessation
An Analysis of Competency Based Activities
To assist smokers in potentially quitting their habit, this paper investigates digital eHealth applications. Based on a dataset provided by research into an eHealth application, it aims to determine if persuasive activities can convince users of the usefulness of competencies determined to assist in quitting smoking. A thematic analysis was used on the participant's qualitative responses to the activities. Using this and the quantitative data derived, insights into the efficacy of persuasion were determined.
It was determined that engagement with the optional qualitative aspect of the data produced similar utility perspectives on the competencies to those who did not comment. It was noted that the general perspective of the competencies rose after completing the activity, however not to a significant degree. Additionally, no notable correlations between age, gender or educational level and increased perception of the competency arose. Several interesting remarks from participants were analysed to offer considerations for any future research in this field. ...
It was determined that engagement with the optional qualitative aspect of the data produced similar utility perspectives on the competencies to those who did not comment. It was noted that the general perspective of the competencies rose after completing the activity, however not to a significant degree. Additionally, no notable correlations between age, gender or educational level and increased perception of the competency arose. Several interesting remarks from participants were analysed to offer considerations for any future research in this field. ...
To assist smokers in potentially quitting their habit, this paper investigates digital eHealth applications. Based on a dataset provided by research into an eHealth application, it aims to determine if persuasive activities can convince users of the usefulness of competencies determined to assist in quitting smoking. A thematic analysis was used on the participant's qualitative responses to the activities. Using this and the quantitative data derived, insights into the efficacy of persuasion were determined.
It was determined that engagement with the optional qualitative aspect of the data produced similar utility perspectives on the competencies to those who did not comment. It was noted that the general perspective of the competencies rose after completing the activity, however not to a significant degree. Additionally, no notable correlations between age, gender or educational level and increased perception of the competency arose. Several interesting remarks from participants were analysed to offer considerations for any future research in this field.
It was determined that engagement with the optional qualitative aspect of the data produced similar utility perspectives on the competencies to those who did not comment. It was noted that the general perspective of the competencies rose after completing the activity, however not to a significant degree. Additionally, no notable correlations between age, gender or educational level and increased perception of the competency arose. Several interesting remarks from participants were analysed to offer considerations for any future research in this field.
Smoking has been one of the great threats to health in recent years, being strongly correlated with multiple negative health consequences, including lung cancer. Recent research suggests that artificial intelligence chatbots can be effective in persuading healthy behavior change. However, these chatbots usually rely on persuasive techniques to achieve their goal. Such techniques depend on identifying and meeting users' needs to be effective. To help improve understanding of the domain of healthy behavior change, we proposed a study which analysed the needs of daily smokers as they emerged from their interactions with a chatbot specifically designed to help them quit. The study performed a thematic analysis on users' free-text feedback, from which a set of 8 themes that directly correspond to 8 different needs were observed. The user needs were correlated with their genders, ages and highest completed education levels. While most of the results indicate that there is no significant correlation between the needs and user characteristics, which suggests that user needs are evenly distributed, certain correlations were highlighted for further analysis.
...
Smoking has been one of the great threats to health in recent years, being strongly correlated with multiple negative health consequences, including lung cancer. Recent research suggests that artificial intelligence chatbots can be effective in persuading healthy behavior change. However, these chatbots usually rely on persuasive techniques to achieve their goal. Such techniques depend on identifying and meeting users' needs to be effective. To help improve understanding of the domain of healthy behavior change, we proposed a study which analysed the needs of daily smokers as they emerged from their interactions with a chatbot specifically designed to help them quit. The study performed a thematic analysis on users' free-text feedback, from which a set of 8 themes that directly correspond to 8 different needs were observed. The user needs were correlated with their genders, ages and highest completed education levels. While most of the results indicate that there is no significant correlation between the needs and user characteristics, which suggests that user needs are evenly distributed, certain correlations were highlighted for further analysis.
Use Reinforcement Learning to Choose Activities for Preparing to Quit Smoking
How Effective a Reinforcement Learning Model is for Choosing Activities that Optimizes the Likelihood that Users Return to the Next Session and the Effort Users Spend on Their Activities?
Unhealthy behaviors such as smoking is the major cause for premature deaths and changing behaviors by oneself can be difficult. That is where eHealth applications come into rescue. One of the recent research explored the possibility of using a Reinforcement Learning model to choose persuasive types for a virtual coach to adopt to persuade people to prepare for smoke-quitting and it has shown advantages. However, there are still more aspects to investigate in this context except the persuasive types of the messages, and this paper intended to further look into using reinforcement learning to choose activities for preparing the users to quit smoking, To be more specific, we implemented and evaluated a reinforcement learning model to choose activities to optimize both the effort spent by the users and also the likelihood of them staying for the next session. The result suggests that reinforcement learning is a promising approach to choose activities for people to prepare for quitting smoking and it can move the users to states that they are more likely to spend a good effort on the activities and are more likely to come back to the next session.
...
Unhealthy behaviors such as smoking is the major cause for premature deaths and changing behaviors by oneself can be difficult. That is where eHealth applications come into rescue. One of the recent research explored the possibility of using a Reinforcement Learning model to choose persuasive types for a virtual coach to adopt to persuade people to prepare for smoke-quitting and it has shown advantages. However, there are still more aspects to investigate in this context except the persuasive types of the messages, and this paper intended to further look into using reinforcement learning to choose activities for preparing the users to quit smoking, To be more specific, we implemented and evaluated a reinforcement learning model to choose activities to optimize both the effort spent by the users and also the likelihood of them staying for the next session. The result suggests that reinforcement learning is a promising approach to choose activities for people to prepare for quitting smoking and it can move the users to states that they are more likely to spend a good effort on the activities and are more likely to come back to the next session.
LLM-PQA
A Natural Language Interface for Machine Learning Model Selection and Training
Many domain experts encounter significant challenges in leveraging machine learning due to the technical complexity of model selection and development. This thesis presents LLM-PQA, a system that enables natural language interaction with machine learning functionalities through large language models. The system combines LLM-based query interpretation with vector search techniques for resource matching and automated machine learning tools for model development. We provide a benchmark dataset for evaluating natural language-driven machine learning systems and conduct experiments examining feature extraction accuracy and embedding technique effectiveness, along with a comprehensive analysis of AutoML tool integration capabilities. Using our benchmark dataset, the evaluation demonstrates that one-shot prompting with LLMs achieves reliable feature extraction, while normalized vector embeddings enable precise resource matching. Analysis of various AutoML tools reveals integration strategies that maintain both technical capability and user accessibility. These findings advance the development of natural language interfaces for machine learning systems while providing insights into effective integration patterns for automated machine learning tools.
...
Many domain experts encounter significant challenges in leveraging machine learning due to the technical complexity of model selection and development. This thesis presents LLM-PQA, a system that enables natural language interaction with machine learning functionalities through large language models. The system combines LLM-based query interpretation with vector search techniques for resource matching and automated machine learning tools for model development. We provide a benchmark dataset for evaluating natural language-driven machine learning systems and conduct experiments examining feature extraction accuracy and embedding technique effectiveness, along with a comprehensive analysis of AutoML tool integration capabilities. Using our benchmark dataset, the evaluation demonstrates that one-shot prompting with LLMs achieves reliable feature extraction, while normalized vector embeddings enable precise resource matching. Analysis of various AutoML tools reveals integration strategies that maintain both technical capability and user accessibility. These findings advance the development of natural language interfaces for machine learning systems while providing insights into effective integration patterns for automated machine learning tools.
Contract review is a critical yet time-consuming process in legal practice, with significant financial implications when errors occur. While Large Language Models (LLMs) have shown promise in legal document processing, they still face challenges with lengthy contracts and complex legal relationships. This research presents an advanced approach to automated contract review by integrating knowledge graphs into Retrieval-Augmented Generation (RAG) frameworks, addressing the limitations of current methodologies.
Through a comprehensive literature review of contract review automation and RAG systems, we conducted systematic experiments comparing RAG approaches with LLMs' in-context learning capabilities. Our empirical analysis validated that RAG-based methods significantly enhance long-context text analysis and information extraction in legal documents, particularly in terms of accuracy and consistency.
Building on these findings, we extensively investigated optimization techniques for the RAG retrieval phase, recognizing its critical role in contract review accuracy. Our experimental evaluation encompassed various chunking strategies, query expansion methods, and re-ranking approaches, establishing best practices for legal document processing.
Our primary contribution is a novel KG-RAG system that enhances contextual understanding in legal document analysis. We evaluate our approach using the Contract Understanding Atticus Dataset (CUAD) and ContractNLI dataset, demonstrating improved performance over traditional RAG implementations and long-context models. The research also explores optimal chunking strategies and investigates the efficiency-effectiveness trade-offs between different model architectures.
Results indicate that our KG-enhanced RAG framework achieves superior performance in identifying and analyzing complex legal relationships while maintaining computational efficiency. The integration of knowledge graphs particularly excels in capturing hierarchical and cross-referential relationships within legal documents, a crucial aspect often overlooked by conventional approaches.
This work advances the field of legal AI by providing a more robust and context-aware approach to contract review, while offering practical insights for implementing AI systems in legal practice. Our findings suggest promising directions for future research in legal document processing, particularly in areas requiring deep contextual understanding and relationship modeling. ...
Through a comprehensive literature review of contract review automation and RAG systems, we conducted systematic experiments comparing RAG approaches with LLMs' in-context learning capabilities. Our empirical analysis validated that RAG-based methods significantly enhance long-context text analysis and information extraction in legal documents, particularly in terms of accuracy and consistency.
Building on these findings, we extensively investigated optimization techniques for the RAG retrieval phase, recognizing its critical role in contract review accuracy. Our experimental evaluation encompassed various chunking strategies, query expansion methods, and re-ranking approaches, establishing best practices for legal document processing.
Our primary contribution is a novel KG-RAG system that enhances contextual understanding in legal document analysis. We evaluate our approach using the Contract Understanding Atticus Dataset (CUAD) and ContractNLI dataset, demonstrating improved performance over traditional RAG implementations and long-context models. The research also explores optimal chunking strategies and investigates the efficiency-effectiveness trade-offs between different model architectures.
Results indicate that our KG-enhanced RAG framework achieves superior performance in identifying and analyzing complex legal relationships while maintaining computational efficiency. The integration of knowledge graphs particularly excels in capturing hierarchical and cross-referential relationships within legal documents, a crucial aspect often overlooked by conventional approaches.
This work advances the field of legal AI by providing a more robust and context-aware approach to contract review, while offering practical insights for implementing AI systems in legal practice. Our findings suggest promising directions for future research in legal document processing, particularly in areas requiring deep contextual understanding and relationship modeling. ...
Contract review is a critical yet time-consuming process in legal practice, with significant financial implications when errors occur. While Large Language Models (LLMs) have shown promise in legal document processing, they still face challenges with lengthy contracts and complex legal relationships. This research presents an advanced approach to automated contract review by integrating knowledge graphs into Retrieval-Augmented Generation (RAG) frameworks, addressing the limitations of current methodologies.
Through a comprehensive literature review of contract review automation and RAG systems, we conducted systematic experiments comparing RAG approaches with LLMs' in-context learning capabilities. Our empirical analysis validated that RAG-based methods significantly enhance long-context text analysis and information extraction in legal documents, particularly in terms of accuracy and consistency.
Building on these findings, we extensively investigated optimization techniques for the RAG retrieval phase, recognizing its critical role in contract review accuracy. Our experimental evaluation encompassed various chunking strategies, query expansion methods, and re-ranking approaches, establishing best practices for legal document processing.
Our primary contribution is a novel KG-RAG system that enhances contextual understanding in legal document analysis. We evaluate our approach using the Contract Understanding Atticus Dataset (CUAD) and ContractNLI dataset, demonstrating improved performance over traditional RAG implementations and long-context models. The research also explores optimal chunking strategies and investigates the efficiency-effectiveness trade-offs between different model architectures.
Results indicate that our KG-enhanced RAG framework achieves superior performance in identifying and analyzing complex legal relationships while maintaining computational efficiency. The integration of knowledge graphs particularly excels in capturing hierarchical and cross-referential relationships within legal documents, a crucial aspect often overlooked by conventional approaches.
This work advances the field of legal AI by providing a more robust and context-aware approach to contract review, while offering practical insights for implementing AI systems in legal practice. Our findings suggest promising directions for future research in legal document processing, particularly in areas requiring deep contextual understanding and relationship modeling.
Through a comprehensive literature review of contract review automation and RAG systems, we conducted systematic experiments comparing RAG approaches with LLMs' in-context learning capabilities. Our empirical analysis validated that RAG-based methods significantly enhance long-context text analysis and information extraction in legal documents, particularly in terms of accuracy and consistency.
Building on these findings, we extensively investigated optimization techniques for the RAG retrieval phase, recognizing its critical role in contract review accuracy. Our experimental evaluation encompassed various chunking strategies, query expansion methods, and re-ranking approaches, establishing best practices for legal document processing.
Our primary contribution is a novel KG-RAG system that enhances contextual understanding in legal document analysis. We evaluate our approach using the Contract Understanding Atticus Dataset (CUAD) and ContractNLI dataset, demonstrating improved performance over traditional RAG implementations and long-context models. The research also explores optimal chunking strategies and investigates the efficiency-effectiveness trade-offs between different model architectures.
Results indicate that our KG-enhanced RAG framework achieves superior performance in identifying and analyzing complex legal relationships while maintaining computational efficiency. The integration of knowledge graphs particularly excels in capturing hierarchical and cross-referential relationships within legal documents, a crucial aspect often overlooked by conventional approaches.
This work advances the field of legal AI by providing a more robust and context-aware approach to contract review, while offering practical insights for implementing AI systems in legal practice. Our findings suggest promising directions for future research in legal document processing, particularly in areas requiring deep contextual understanding and relationship modeling.
Immersive technologies, such as head-mounted devices for Virtual Reality (VR), have experienced increased availability on the consumer market. This opens up numerous possibilities for scientific research and industrial applications. But, configuring and combining data from various sources to create an effective visualization to explore, collaborate in and interact with, has only been researched in limited aspects. This thesis aims to demonstrate the feasibility of integrating these diverse aspects into a single application while preserving their individual potential.
To achieve this objective, an extensive analysis of the literature was done to establish the state of the art. All aspects necessary to create a single application were investigated to define how much of the literature fully or partly implemented them. Findings of the literature were summarized through definition, progress, limitations and improvements.
Based on the analysis, a proof of concept was designed and implemented, which established a virtual reality framework for real-time analysis and exploration of heterogeneous urban data. The framework involved the development of both a client and server applications with real-time communication. The server application processes and caches input data based on a structured configuration file, while the client application renders and visualizes points and shapes based on instructions from the server. The layered processing approach in this proof of concept facilitated real-time data handling and exploration. Which also makes it possible to collaborate with multiple people working on or manipulating the same data at the same time.
Evaluation of the server application showed its capability of processing over 1.000.000 objects from GeoJSON data. After initial rendering, the client application could update over 4.000 objects with a latency of just 200ms, which is considered to be interactive. Feedback from experts further supported the validation of the proof of concept application as a viable solution to the research question.
This research established the validity of a configurable real-time analysis and exploration solution in virtual reality. With further development and research, this tool holds potential for being used in diverse environments dealing with urban data. ...
To achieve this objective, an extensive analysis of the literature was done to establish the state of the art. All aspects necessary to create a single application were investigated to define how much of the literature fully or partly implemented them. Findings of the literature were summarized through definition, progress, limitations and improvements.
Based on the analysis, a proof of concept was designed and implemented, which established a virtual reality framework for real-time analysis and exploration of heterogeneous urban data. The framework involved the development of both a client and server applications with real-time communication. The server application processes and caches input data based on a structured configuration file, while the client application renders and visualizes points and shapes based on instructions from the server. The layered processing approach in this proof of concept facilitated real-time data handling and exploration. Which also makes it possible to collaborate with multiple people working on or manipulating the same data at the same time.
Evaluation of the server application showed its capability of processing over 1.000.000 objects from GeoJSON data. After initial rendering, the client application could update over 4.000 objects with a latency of just 200ms, which is considered to be interactive. Feedback from experts further supported the validation of the proof of concept application as a viable solution to the research question.
This research established the validity of a configurable real-time analysis and exploration solution in virtual reality. With further development and research, this tool holds potential for being used in diverse environments dealing with urban data. ...
Immersive technologies, such as head-mounted devices for Virtual Reality (VR), have experienced increased availability on the consumer market. This opens up numerous possibilities for scientific research and industrial applications. But, configuring and combining data from various sources to create an effective visualization to explore, collaborate in and interact with, has only been researched in limited aspects. This thesis aims to demonstrate the feasibility of integrating these diverse aspects into a single application while preserving their individual potential.
To achieve this objective, an extensive analysis of the literature was done to establish the state of the art. All aspects necessary to create a single application were investigated to define how much of the literature fully or partly implemented them. Findings of the literature were summarized through definition, progress, limitations and improvements.
Based on the analysis, a proof of concept was designed and implemented, which established a virtual reality framework for real-time analysis and exploration of heterogeneous urban data. The framework involved the development of both a client and server applications with real-time communication. The server application processes and caches input data based on a structured configuration file, while the client application renders and visualizes points and shapes based on instructions from the server. The layered processing approach in this proof of concept facilitated real-time data handling and exploration. Which also makes it possible to collaborate with multiple people working on or manipulating the same data at the same time.
Evaluation of the server application showed its capability of processing over 1.000.000 objects from GeoJSON data. After initial rendering, the client application could update over 4.000 objects with a latency of just 200ms, which is considered to be interactive. Feedback from experts further supported the validation of the proof of concept application as a viable solution to the research question.
This research established the validity of a configurable real-time analysis and exploration solution in virtual reality. With further development and research, this tool holds potential for being used in diverse environments dealing with urban data.
To achieve this objective, an extensive analysis of the literature was done to establish the state of the art. All aspects necessary to create a single application were investigated to define how much of the literature fully or partly implemented them. Findings of the literature were summarized through definition, progress, limitations and improvements.
Based on the analysis, a proof of concept was designed and implemented, which established a virtual reality framework for real-time analysis and exploration of heterogeneous urban data. The framework involved the development of both a client and server applications with real-time communication. The server application processes and caches input data based on a structured configuration file, while the client application renders and visualizes points and shapes based on instructions from the server. The layered processing approach in this proof of concept facilitated real-time data handling and exploration. Which also makes it possible to collaborate with multiple people working on or manipulating the same data at the same time.
Evaluation of the server application showed its capability of processing over 1.000.000 objects from GeoJSON data. After initial rendering, the client application could update over 4.000 objects with a latency of just 200ms, which is considered to be interactive. Feedback from experts further supported the validation of the proof of concept application as a viable solution to the research question.
This research established the validity of a configurable real-time analysis and exploration solution in virtual reality. With further development and research, this tool holds potential for being used in diverse environments dealing with urban data.
This thesis aims to assess robustness of networks by evaluating the performance of node attack strategies, the applicability and accuracy of different approaches, and to increase robustness of networks through analysing protecting methods, including link addition and node protection strategies. To be specific, the relative size of the Largest Connected Component (rLCC) and Average Two-Terminal Reliability (ATTR) are chosen to be the performance metrics to evaluate the robustness of networks. In the assessment of network robustness, simulations are employed for ten distinct attack strategies, which include random attack, non-updated and updated degree attacks, non-updated and updated stochastic degree attacks, non-updated and updated betweenness attacks, and greedy attack. Three analytical approximation methods are implemented and discussed, alongside the testing of a machine learning-based approach. Regarding the enhancement of network robustness, the comparative analysis involves thirteen link addition strategies and four node protection strategies. In all node removal scenarios, it is observed that the updated betweenness attack emerges as the most harmful strategy. Moreover, analytical approximations are proved to be effective means of evaluating network robustness in the scenario of predicting robustness for synthetic networks under random attack. Machine learning based methodologies show high accuracy in predicting robustness of synthetic networks, exhibiting acceptable error rates in predicting robustness of real-world networks. Besides, the updated betweenness link addition strategy and targeted betweenness-based node protection exhibit the highest efficacy in protecting networks against the most harmful attacks.
...
This thesis aims to assess robustness of networks by evaluating the performance of node attack strategies, the applicability and accuracy of different approaches, and to increase robustness of networks through analysing protecting methods, including link addition and node protection strategies. To be specific, the relative size of the Largest Connected Component (rLCC) and Average Two-Terminal Reliability (ATTR) are chosen to be the performance metrics to evaluate the robustness of networks. In the assessment of network robustness, simulations are employed for ten distinct attack strategies, which include random attack, non-updated and updated degree attacks, non-updated and updated stochastic degree attacks, non-updated and updated betweenness attacks, and greedy attack. Three analytical approximation methods are implemented and discussed, alongside the testing of a machine learning-based approach. Regarding the enhancement of network robustness, the comparative analysis involves thirteen link addition strategies and four node protection strategies. In all node removal scenarios, it is observed that the updated betweenness attack emerges as the most harmful strategy. Moreover, analytical approximations are proved to be effective means of evaluating network robustness in the scenario of predicting robustness for synthetic networks under random attack. Machine learning based methodologies show high accuracy in predicting robustness of synthetic networks, exhibiting acceptable error rates in predicting robustness of real-world networks. Besides, the updated betweenness link addition strategy and targeted betweenness-based node protection exhibit the highest efficacy in protecting networks against the most harmful attacks.
In real-world scenarios, users provide invaluable data; however, this data is inherently incoherent, incomplete, and duplicated, i.e., different data rows refer to the same real-world object. Merging duplications to a single entry broadens the knowledge of a given real-world object represented within a data set. Applying a straightforward cross-join operation to check for duplications is infeasible and intractable in big data scenarios. Instead of comparing all rows, the similarity can be approximated by Locality-Sensitive Hashing (LSH). LSH’s MinHash creates colliding hashes on common tokens. With enough matching hashes, two rows can be deemed similar. The applications of LSH within an Entity Resolution (ER) pipeline on incoherent and incomplete big data are thoroughly investigated in this work. An ER pipeline is introduced to make the deduplication problem tractable. This pipeline consists of four phases: preprocessing, blocking, matching, and clustering. In the preprocessing, the data is made coherent. In the blocking, the data is divided into smaller main blocks by blocking on complete properties. These blocks contain a majority of True-Matches. The created blocks can be augmented with additional blocks that rely on incomplete properties. In the matching phase, only the Matches found within the blocks are compared, significantly reducing the required number of comparisons. The incomplete properties cannot be transformed into features for every Match; therefore, classifiers are trained per property combination, allowing for optimal Match classification. In clustering, the duplicated rows are merged into a single Entity. These Entities are created with disconnected sub-graphs of Positive-Matches. These sub-graphs are particularly susceptible to noise. Therefore, novel implementations of MinHash are used for noise removal and finding missed True-Matches. The created Entities were tested on an existingmodel. Blocks constructed with MinHash on N-gram tokens allow spelling mistakes to be overcome within the user data, which improves the Pairs Completeness (PC) at the cost of Pairs Quality (PQ) and Reduction Ratio (RR). Introducing augmentative blocks increases the PC significantly at the cost of PQ and a low cost of RR. Multiple runs are investigated throughout the pipeline. The runs relying on MinHash showed increased PC at the cost of PQ and RR. A novel concept of a double-pass pipeline is introduced, in which the baseline ER pipeline should be initially applied, followed by a more expensive error-correctiveMinHash ER pipeline that can be used on Entities created by the first pass, combining the strengths of the investigated runs.
...
In real-world scenarios, users provide invaluable data; however, this data is inherently incoherent, incomplete, and duplicated, i.e., different data rows refer to the same real-world object. Merging duplications to a single entry broadens the knowledge of a given real-world object represented within a data set. Applying a straightforward cross-join operation to check for duplications is infeasible and intractable in big data scenarios. Instead of comparing all rows, the similarity can be approximated by Locality-Sensitive Hashing (LSH). LSH’s MinHash creates colliding hashes on common tokens. With enough matching hashes, two rows can be deemed similar. The applications of LSH within an Entity Resolution (ER) pipeline on incoherent and incomplete big data are thoroughly investigated in this work. An ER pipeline is introduced to make the deduplication problem tractable. This pipeline consists of four phases: preprocessing, blocking, matching, and clustering. In the preprocessing, the data is made coherent. In the blocking, the data is divided into smaller main blocks by blocking on complete properties. These blocks contain a majority of True-Matches. The created blocks can be augmented with additional blocks that rely on incomplete properties. In the matching phase, only the Matches found within the blocks are compared, significantly reducing the required number of comparisons. The incomplete properties cannot be transformed into features for every Match; therefore, classifiers are trained per property combination, allowing for optimal Match classification. In clustering, the duplicated rows are merged into a single Entity. These Entities are created with disconnected sub-graphs of Positive-Matches. These sub-graphs are particularly susceptible to noise. Therefore, novel implementations of MinHash are used for noise removal and finding missed True-Matches. The created Entities were tested on an existingmodel. Blocks constructed with MinHash on N-gram tokens allow spelling mistakes to be overcome within the user data, which improves the Pairs Completeness (PC) at the cost of Pairs Quality (PQ) and Reduction Ratio (RR). Introducing augmentative blocks increases the PC significantly at the cost of PQ and a low cost of RR. Multiple runs are investigated throughout the pipeline. The runs relying on MinHash showed increased PC at the cost of PQ and RR. A novel concept of a double-pass pipeline is introduced, in which the baseline ER pipeline should be initially applied, followed by a more expensive error-correctiveMinHash ER pipeline that can be used on Entities created by the first pass, combining the strengths of the investigated runs.
This thesis’ research concerns the time-dynamics of a complex geographical network of municipalities, i.e. the Dutch Municipality Network over the period 1830-2019. By analysing 190 years of socioeconomic statistical data and applying contemporary tooling from network science and geographic information systems (GIS), the findings from this research can provide a new approach and supportive methods for policymakers, statistical offices, researchers and businesses (to decide when and where to invest).
...
This thesis’ research concerns the time-dynamics of a complex geographical network of municipalities, i.e. the Dutch Municipality Network over the period 1830-2019. By analysing 190 years of socioeconomic statistical data and applying contemporary tooling from network science and geographic information systems (GIS), the findings from this research can provide a new approach and supportive methods for policymakers, statistical offices, researchers and businesses (to decide when and where to invest).