With the increasing demand for artificial intelligence (AI), intelligent systems have become deeply integrated into various aspects of modern life, including autonomous driving, smart assistants on mobile devices, and powerful online language models such as ChatGPT. In addition, the emergence of generative models for text and image synthesis has significantly reduced the cost of ac cessing and interacting with information. However, the advancement of these AI applications comes at the cost of ever-growing computational and memory requirements. These requirements pose substantial challenges even for high end computing systems, and become prohibitive when deploying AI models on resource-constrained platforms such as embedded devices and Internet of Things (IoT) nodes. This thesis presents a distributed inference framework for Transformer mod els where we design a fine-grained, channel-wise parameter partitioning scheme. Importantly, the implementation of our framework is also independent of con ventional AI frameworks such as PyTorch, making it efficient, portable, and adaptable to virtually any compute-capable device. We begin by analyzing the computational and memory limitations of main stream hardware platforms, highlighting the motivation to aggregate multiple low-power devices to collectively execute AI workloads. Through software based simulation, we validate the correctness of the partitioned inference scheme and demonstrate that it introduces no functional deviation from unpartitioned single-device execution. The simulation also facilitates precise estimation of data flow and compute demand across multiple collaborating devices. Furthermore, we introduce a full-stack load balancing algorithm that enables adaptive task allocation based on heterogeneous hardware specifications, taking into account factors such as bandwidth, memory capacity, and communication latency. In summary, this thesis proposes a split, practical, and high-granularity Trans former inference framework that is compatible with heterogeneous hardware configurations, offering a promising step toward enabling AI inference distributively on a network of resource-constrained embedded platforms.

There has been a steady increase in technologies that leverage Deep Learning (DL) techniques on resource-constrained devices for real-time processing. While DL techniques are adept at recognition tasks, their performance depends on the training process. Training data is seldom fully representative of the deployment scenario, requiring retraining to preserve accuracy. Many works have proposed on-device learning techniques that enable training DL techniques like Convolutional Neural Networks (CNNs) on microcontroller units (MCUs), without requiring data to be sent to the cloud. These methods have yielded promising accuracy improvements while training with low memory footprints. However, limited research has been done on the energy implications of doing so.

This thesis presents methods for energy-aware on-device learning on MCUs. It leverages the principle of updating specific layers of a CNN, proposed in past works, to fit memory constraints. We propose an energy-accuracy trade-off objective based on computational costs (in MACs) and accuracy improvement to select which layers to train. Furthermore, we demonstrate how computationally light search algorithms can adequately maximize the newly defined objective for layer selection. Evaluations show that our approach saves up to 200mJ of energy on-device while yielding simulation accuracies similar to a recent study under the same conditions.

Online databases contain extensive collections of (bio)chemical reactions serving as valuable resources for a variety of applications. However, these large datasets often suffer from incomplete reaction data missing, for example, co-reactants and by-products. Machine learning can help to predict these missing molecules in partial reactions. In this study, we adapt an existing transformer model to enhance its capability in completing these incomplete reactions. We retrain the model using a more diverse dataset of atom-balanced ground truth reactions and introduce both soft and hard atom-balance constraints to improve the completeness and chemical validity of the predictions. Our findings indicate that models trained with soft constraints in their loss function do not demonstrate improved balancing performance and require further tuning. Conversely, the implementation of hard atom-balance constraints during constrained beam search, where we restrict predicting tokens that violate the atom-balance of the prediction, effectively improves the performance of transformer-based models in reaction completion tasks. However, this approach also presents the risk of inaccurately balancing reactions; a limitation that is difficult to identify without chemical expertise, underscoring the necessity for reliable ground truth data to evaluate the predictions.

The Cytomegalovirus (CMV) serostatus, of both donor and patient, plays an important role in allogeneic hematopoietic stem cell transplantation, yet it is only known for 19% of donors in the global database. In this research, the other available data in the global database of the World Marrow Donor Association will be used to predict CMV serostatus for donors who’s status is unknown. In a statistical analysis, features such as sex, registry, ethnicity, age and height were found to be informative. A particular focus was put on investigating the relation between Human Leukocyte Antigen (HLA) and CMV. Previous literature on this relation consists of studies on small cohorts, Machida et al., 1998 (N=125) and Hassan et al., 2016 (N=1 849). The large global database (N=8 707 407 allows us to evaluate this relation for many more different HLA groups and validate their findings using a meta-analysis across registries. In this analysis, the majority of HLA groups showed small effects on the CMV serostatus. However, a few groups showed consistent large increases in likelihood to be CMV seropositive. This indicates that there is a biological relation between specific HLA and CMV serostatus. To predict CMV serostatus, a simple logistic regression classifier was iterated on by adding features and using more complex models. The best performing classifier uses XGBoost on all features in the database. This classifier has an AUC performance of 0.70. Although this AUC is too low to rely on it for patient-donor matching, it does show non-trivial predictive power. The classifier could be further improved by using a better embedding for the HLA that includes the similarities between different HLA alleles.

This study examines dataset annotation practices in influential NeurIPS research. Datasets employed in highly cited NeurIPS papers were assessed based on criteria concerning their item population, labelling schema, and annotation process. While high-level information, such as the presence of human labellers and item population, is present in most cases, procedural details of the annotation process are poorly reported. Notably, 48% of datasets lack details on annotator training, 43% omit inter-rater reliability, and 28% are not publicly accessible. Temporal comparisons show minor improvements, but no substantial progress in reporting annotation methodology. A complementary analysis of 49 NeurIPS papers published since 2020 shows that researchers often discuss the broader impact of their work, yet do not include datasets or their annotations in these assessments. These findings highlight a lack of standardisation in annotation reporting and call for more robust practices that ensure transparency, auditability, and reproducibility in machine learning research.

The output of machine learning (ML) models can be only as good as the data that is fed into them. Because of this, when making datasets for creating ML models, it is important to ensure the quality of the data. This is especially true of human labeled data, which can be hard to standardize and assess the quality of. To assess the annotation practices of human labeled data in the field of machine learning, this paper investigates the datasets used in the highest cited papers in the AAAI Conference on Artificial Intelligence, an influential machine learning conference. After extracting the datasets from 75 papers in three overlapping publication periods, the top 20 datasets were evaluated from each period. The results showed that the majority of datasets do not use or underreport significant annotation practices, specifically about the annotators and the annotation process. This raises concern for the conference and the field more broadly, as the most influential papers build their machine learning algorithms on quite possibly low quality data. However, there is some hope for the field in this regard as the more recent papers use datasets with better quality annotation practices.

This study investigates annotation and reporting practices in machine learning (ML) research, focusing on societally impactful applications presented at the IEEE/CVF Computer Vision and Pattern Recognition (CVPR) conferences. By structurally analyzing the 75 most-cited CVPR papers from the past 2, 5, and 15 years, we evaluate how the human annotations foundation of supervised ML is documented. We introduce a 27-field annotation-reporting schema and apply it to 60 datasets, revealing that nearly 30% of relevant information is routinely omitted. Key findings include the pervasive underreporting of annotator details such as training, prescreening, and inter-rater reliability (IRR) metrics. While popular datasets like COCO and ImageNet exhibit widespread use, transparency about annotation methodologies remains inconsistent. The impact of a few fields shows that basic metadata, such as the selection process of annotators and how the labels' overlap is managed, strongly anticipate overall documentation quality. Our findings support previous calls for standardization and underscore the need for institutionalized reporting practices to ensure reproducibility, fairness, and trust in ML systems.

High-impact vision research still rests on datasets whose labels arrive via opaque, rarely documented pipelines. To understand how serious the problem is inside a large venue, we audited 75 TPAMI papers (2009-2024) that rely or introduce datasets. Each dataset was coded against a 27-item checklist adapted from Garbage in, Garbage out, spanning annotator recruitment, training, compensation, overlap-resolution and more. Across the corpus, 37% of the expected annotation metadata is missing; the rate changes little between recent (2022-24) and older cohorts. The scarcest fields are labeller-population rationale (76.6% absent), prescreening criteria (73.4%), total annotators (68.8%), compensation (67.2%) and training procedures (62.5%). Documentation quality shows virtually no correlation with a paper’s citation impact, suggesting community prestige does not buy transparency. A handful of well—curated datasets achieve >75% completeness, proving that thorough documentation is possible when incentives align. The median TPAMI benchmark still ships with an unverifiable "ground truth", threatening the reproducibility and fairness claims of downstream models. We advocate that journals and conferences require a concise, checklist-based annotation statement, mirroring existing ethics and reproducibility forms, to ensure future vision systems are built (and evaluated) on transparent, trustworthy data foundations.

This study gives an overview of the data collection and annotation practices of the datasets used by the most impactful papers published by the Association of Computational Linguistics (ACL). This was achieved by selecting the most highly cited papers published within the ACL anthology across 3 periods (published in the past 2, 5 and 15 years). Afterwards, the datasets used by those papers were extracted and filtered to retain the most impactful ones. Finally, a carefully crafted annotation schema was used to find out information regarding key aspects of the datasets in order to qualitatively analyze them. As a result of this analysis, it was first found that (1) there are fewer datasets used on average in the past 2 years and that there is little overlap with the datasets used by papers published in the past 5 or 15 years. (2) Secondly, there are various concerns related to those key aspects, such as the relatively high (∼36%) and unregulated use of the Amazon Mechanical Turk crowdsourcing platform for the construction of datasets. Another concern is information frequently missing about any rationale regarding labeller population, prescreening, inter-rater reliability and rationale regarding sample size - missing ∼77%, ∼63%, ∼19-56%, and ∼81% of the time. However, reporting practices for most of those issues have slightly improved within datasets used in the past 2 years. (3) Finally, around one third of the information sought was missing across all periods. However, the state of the domain has been generally improving, with a lower one fourth of the information missing from datasets used in the past 2 years. Some recommendations are given in order to overcome those challenges, the most important of which being that each academic organization should require their submissions to include a reporting template in their papers.

This thesis investigates reducing carbon emissions in code generation using large language models (LLMs) by comparing function-level and line-level code completions across models of different sizes (1.5B and 9B parameters). The study utilises the BigCodeBench dataset, comprising 1,140 Python programming problems, to evaluate the energy consumption, test accuracy, and time efficiency of code completions. The models, 4-bit quantised and run on a CPU, performed 30 function-level completions and 30 line-level completions for each line, which were tested for correctness. Results indicate that, while line-level completions require slightly more energy per token, they are more efficient overall in terms of total energy consumption and token usage. The smaller model with line-level completions showed significant reductions in carbon emissions, achieving an average tenfold reduction compared to the large model with function-level completions. With the large model, line-level completions achieved a $4.5\times$ reduction in carbon emissions compared to function-level completions. Line-level completions were more token-efficient, wasting less than 1\% of energy, compared to 20\% for function-level completions. From a sustainability perspective, line-level completions offer a practical strategy to reduce the environmental impact of code generation tasks while maintaining strong performance. The study suggests that optimising completion strategies could help balance energy consumption, test accuracy, and time efficiency. Future research could explore a broader range of model sizes, fine-tuning models specifically for line-level completions, a performance decrease in solution length, and alternative validation metrics to assess code generation performance.

The rapid rise in the popularity of large language models has highlighted the need for extensive datasets, especially for training on code. However, this growth has also raised important questions about the legal implications of using code in large language model training, particularly regarding the potential infringement of code licenses. At the same time, the availability of clean datasets for evaluating these models is becoming increasingly limited, due to a high risk of contamination which restricts the capacity for reliable research. On top of that, this requires researchers to repeatedly perform data curation steps in order to evaluate their models on downstream tasks, based on previously unseen data. This process is not only time- and resource-intensive but also introduces potential inconsistencies across studies, which can impact their reproducibility.
We address these challenges through a comprehensive licensing analysis and by developing robust datasets to support accurate and reproducible large language model evaluations. We compiled a list of 53 large language models trained on file-level code and analyzed their datasets, discovering pervasive license inconsistencies despite careful selection based on repository licenses. Our analysis, covering 514M code files, reveals 38M exact duplicates of strong copyleft code, and 171M file-leading comments, 16M of which are under copyleft licenses and another 11M discouraging unauthorized copying. To further understand the depth of non-permissive code in public training datasets, we developed StackLessV2, a strong copyleft Java dataset decontaminated against The Stack V2 to facilitate accurate model evaluations. Our results revealed that non-permissive code is also present at the near-duplication level, although, this represents a gray area in terms of legal interpretation, where the boundary between acceptable reuse and license violation is still unclear, emphasizing the need for further legal clarification. Finally, we extend on this and introduce The Heap, a large multilingual copyleft dataset covering 57 programming languages, specifically deduplicated to avoid contamination from existing open training datasets. The Heap offers a solution for conducting fair, reproducible evaluations of large language models without the significant overhead of the data curation process.

This thesis investigates the occurrence and mitigation of Sequential Social Dilemmas (SSDs) in Local Energy Communities (LECs) managed through Multi-agent Reinforcement Learning (MARL). LECs have great potential as pivotal elements in the green energy transition, yet the inherent conflict between individual incentives and community-wide objectives creates SSD scenarios that challenge learning processes. To address these issues, we propose an agent-centric approach and develop a custom MARL environment where agents interact via a communal battery system and a local trading mechanism.

We systematically investigate the impact of resource constraints and social interactions on the agents' learning. In non-cooperative settings, limited resources impede policy optimization, while the introduction of a shared battery reveals SSD dynamics driven by both greed and fear factors. Our experiments show that rescaling the training data leads agents to adopt more cooperative behaviors, and that reward function modifications incentivizing community-friendly battery use cause a significant increase in social welfare. These mitigation techniques are further validated in a realistic LEC environment with multiple, heterogeneous households engaging in trading and storage actions.

The contributions of this thesis are threefold: (1) the proposal of a new agent-centric MARL environment for LECs, (2) the demonstration of SSDs impacting MARL performance in these decentralized energy systems, and (3) the introduction of concrete strategies for aligning individual and community incentives.

Bayesian Networks (BNs) are widely utilized across various industrial sectors to optimize processes, with an emerging focus on the collaboration across multiple parties. While most realistic scenarios require handling a mixture of categorical and continuous data simultaneously, the current state-of-the-art only supports collaborative inference on purely discrete models. The Junction Tree enables efficient and accurate inference on hybrid models but has not been implemented for confidential scenarios yet. To address this gap, we introduce Hybrid CCJT, an innovative framework for confidential multiparty inference in hybrid domains, offering: (i) a method to construct a collaborative, strongly-rooted junction tree for efficient and secure inference, (ii) a confidential-
preserving inference protocol for Hybrid BNs, (iii) an optimized message-passing scheme that
improves communication efficiency even in the purely discrete domain. Our extensive evaluation
show that Hybrid CCJT improves the predictive accuracy of continuous target variables by an average of 32% in Mean Squared Error and reduce the communication cost up to 86-fold, against the best state-of-the-art baseline.

Recommender systems are widely used in modern lives and contribute to many industries. Therefore, methods to evaluate and improve them are important. Nowadays, much research has been done to improve the system aspects such as algorithms. However, user experiences are not only affected by the systems but heavily rely on the context when using the systems. Therefore, the research on user aspects to understand their experiences is as important. This study contributes an approach that uses collaborative reflection to find insights into users' experiences with recommender systems. Using this approach, this study presents the influences of context on user experiences with recommender systems. This study investigates the importance of situational and personal contexts like mood, time, and location in shaping user satisfaction with recommendations. The research adopts a method based on collaborative reflection, where participants engage in tasks using their YouTube watch history, paired with another individual for real-time discussion. By analyzing contextual influences and the values users wish to achieve, the study identifies key patterns in user behavior and insights into personal preferences. This research not only contributes to the evaluation of recommender systems but also highlights the need for systems to align with both the goals of users and broader societal values. The usability of the proposed method was tested to be successful in crowdsourcing, yielding practical implications for future evaluation and improvements of recommender systems.

In the field of public transportation, environmentally friendly and convenient transportation modes are the future trends. The ride-hailing services is an important component of them. However, current ride-hailing systems, particularly the matching systems, still have issues related to low system efficiency and bad user experience. Although existing ride-hailing rider-driver matching system can allocate travel demands and drivers to a certain extent, they still have deficiencies in certain scenarios. For example, they cannot ensure effective rider-driver matching during peak hours, or they cannot find a good balance between pick-up distance and matching rate. As Reinforcement learning (RL) has been proven in many studies to be applicable and effective in solving complex and dynamic optimization problems. This study aims to explore how Reinforcement Learning (RL) can be adapted to the ride-hailing matching system to optimize system efficiency and user experience through a dynamic matching radius policy. The research objective of this study is to simulate an actual ride-hailing system and use RL to train a policy. This policy can output an optimized dynamic matching radius in real-time based on real-time rider-driver demand-supply relationship, hence achieving a higher matching rate, a shorter average pick-up distance, and a higher driver utilization rate of the ride-hailing system.
Adapting Reinforcement Learning (RL) to optimize the ride-hailing system's matching radius has several difficulties and challenges due to the uncertainties in the real-world rider-hailing market. Traditional approaches are normally static, solving the matching problem at specific times through mathematical models. However, these methods often perform inconsistently when dealing with fluctuating ride-hailing supply-demand relationships, particularly during peak hours. On the other hand, the dynamics and complexity of the ride-hailing market and the ride-hailing environment also make it difficult to model the ride-hailing system. The ride-hailing market is easily affected by many variables, such as weather conditions and local traffic conditions. When quantitatively optimizing the matching radius of the ride-hailing matching system, it is critical to reasonably control irrelevant variables. To address these challenges, this study models the ride-hailing matching problem as a Markov Decision Process (MDP). Based on the defined MDP, a ride-hailing matching simulator is developed. Some assumptions and simplifications are also made to ensure high realism while reasonably controlling irrelevant variables and uncertainties. Multi-replay-buffer Deep Deterministic Policy Gradient (MDDPG) algorithm is then applied to handle the optimization problem of the ride-hailing matching radius. Through the interactions between the MDDPG agent and the developed simulator, feedback rewards are received for the agent to improve the policy. The proposed method is then validated in a case study showcasing the application of the developed simulator and the RL algorithm in a real-world scenario in Austin, Texas. The case study includes an analysis of the current ride-hailing market in Austin, how to apply the simulator based on it, the implementation details of the RL algorithm, and the resulting performance improvements. The results of the case study show that the actions obtained from the proposed method outperform all the baselines in multiple scenarios, highlighting the benefits of using Reinforcement Learning to improve ride-hailing efficiency and user experience.
To conclude, the optimization method proposed in this study applies an advanced Reinforcement Learning approach to the ride-hailing system, successfully improving overall efficiency and user experience. The results of this research demonstrate the potential of Reinforcement Learning in optimizing ride-hailing matching systems, offering a promising direction for further exploration. This study lays a solid foundation for future research to build upon, encouraging the development of more optimization methods with RL technologies that can enhance the effectiveness and adaptability of ride-hailing system in increasingly complex and dynamic environments.

Central banks communicate their monetary policy plans to the public through meeting minutes or transcripts. These communications can have immense effects on markets and are often the subjects of studies in the financial literature. The recent advancements in Natural Language Processing have prompted researchers to analyze these communications using Transformer-based Large Language Model (LLM) classifiers. The use of LLMs in finance and other high-stakes domains calls for a high level of trustworthiness and explainability of those models. We focus on Counterfactual Explanations, a form of Explainable AI that explains a model's classification by proposing an alternative to the original input. We use three types of CE generators for LLM classifiers on a recent dataset consisting of sentences taken from FOMC communications to assess the usability of their explanations. We perform three experiments comparing different types of generators, one using a selection of quantitative metrics and two involving human evaluators, including central bank employees. Our findings suggest that non-expert and expert evaluators prefer counterfactual methods that apply minimal changes to the texts; however, the methods we analyze might not handle the domain-specific vocabulary well enough to generate plausible explanations for our task. We discuss shortcomings in the choice of evaluation metrics in the literature on text CE generators and propose refined definitions of the fluency and plausibility qualitative metrics.

Machine Learning (ML) algorithms have the potential to reproduce biases that already exist in society, a fact that leads to scholarly work trying to quantify algorithmic discrimination through fairness metrics. Although there are now a plethora of metrics, some of them are even contradictory, making fairness become a problem of knowing which measurement to choose over another. Consequently, scholars began considering that fairness should be discussed by placing algorithms in their social contexts. Since (1) these social aspects are related to structures of discrimination and (2) feminism aims to criticise discrimination against the marginalised, I introduce the possibility of analysing the social context of ML algorithms through a feminist lens. By doing this, I highlight social and political aspects that are equally important to consider for a faithful discussion on fairness: corporate lobbying, the lack of diverse hiring which leads to fairness discussions that do not consider the experiences of marginalised groups and, lastly, the broader context that an algorithm is used in. Moreover, I emphasise how feminist ethics of care constitute an essential framework for a conversation about actually implementable fairness solutions, since it shows the need to listen to both the marginalised community and to the developers who might want to build fairer ML but currently cannot. Having built a bridge between the hegemony and the feminist camp, I highlight how Northpointe’s (now Equivant’s) Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) algorithm can be considered biased against black people. Through this, I illustrate how feminist considerations bring clarity to fairness debates by helping choose a fairness metric or by claiming that an algorithm is unfair by nature and should be abolished. To follow, I use the same feminist critiques to draw attention to the possible weak points of current sociotechnical solutions. For instance, the EU AI Act risks being too susceptible to company lobbying, leading to not strict enough regulations. Furthermore, the AI committee should ensure that they hire a diverse group of people in order to develop regulations that positively consider all marginalised groups. Lastly, I highlight how ethics education is essential for creating a new generation of responsible engineers. Considering this, I emphasise the urgency of making ethics courses at TU Delft (and not only) more interdisciplinary by interacting more with critique points coming from the social sciences. This will open up possibilities for more research tackling fairness from a multitude of perspectives.

This project aims to harness the potential of music-making and robotic interaction to enhance creative expression and cognitive function among individuals with cognitive impairment and dementia. With the aging population, there is a growing demand for innovative interventions that support cognitive health and active engagement. Music therapy has demonstrated effectiveness in stimulating cognitive function and emotional expression in individuals with dementia. Creative expression through music serves as a unique outlet, fostering cognitive functions and emotional well-being. This study explores the synergy between music therapy and Socially Assistive Robots (SARs) to develop a more immersive therapeutic intervention. It represents an exploratory investigation into an end-to-end robotic intervention, proposing various interaction elements and examining their functionalities. Each element is designed to foster engagement, enhance perceptions of collaboration, and promote feelings of creativity.
In this thesis, we propose an end-to-end interactive music-making experience designed for use with the Pepper robot, an SAR. The system features a user-friendly interface with eight color-coded boxes, each corresponding to a musical note. Users simply tap the boxes to create melodies. The Pepper robot acts as a guide, assisting users in interacting with the interface. It additionally implements an engagement tracking system by monitoring user interaction through the screen taps on the interface and provides real-time feedback and encouragement. If a period of inactivity is detected, Pepper gently nudges the user to re-engage. Furthermore, the robot functions as a collaborative musical partner, providing rhythmic accompaniment if the user desires. The system also records user-created music and provides playback functionality, allowing users to revisit their compositions.
Methodologically, the study involves an end-to-end system comprising an intelligent music-making interface and an interactive robot providing real-time feedback and rhythmic accompaniment. Insights from the exploratory study highlight the benefits of real-time feedback in enhancing engagement, particularly among participants with musical backgrounds. However, rhythmic accompaniment shows mixed results in fostering collaboration, indicating a preference among participants for emotional connection in collaborative settings. Since this is an exploratory study, the empirical study focuses on healthy older adults, a population with an increased risk of cognitive decline. This is because individuals with dementia are a vulnerable group. Music interventions have shown promise in improving cognitive function and engagement in individuals with dementia. Therefore, this study informs the design of future interventions for people with early-stage dementia.\\
Key findings underscore the potential of real-time feedback and interaction in promoting engagement in the activity. The intelligent music interface also shows potential to support creative exploration, albeit with improvements needed for advanced musical participants. Participants appreciate the playback feature, enhancing their sense of creative ownership and motivation. Despite promising outcomes, the study acknowledges limitations in sample size and participant demographics, primarily recruiting from music-engaged older adults rather than the target demographic of individuals with cognitive impairments.
Future research directions include expanding participant diversity, refining robot interaction capabilities, and addressing technical challenges to improve system usability and accessibility. Integrating findings from ongoing research on music and memory could further enhance personalized interventions. Ultimately, this study lays the groundwork for future developments in robotic interventions that promote well-being through music therapy for individuals with cognitive impairments.

UTURN aims to maximize the matching rate on its freight transport platform by efficiently connecting shippers with suitable carriers. To support this matching process, UTURN required a solution that was additive rather than restrictive on the platform. To achieve this, our research leverages recommendation systems and predictive models to help shipments find appropriate carriers in time while avoiding sending unnecessary email recommendations. We addressed three main questions: how to maximize matching probability through recommendations, how to balance the recommendation frequency to prevent spam, and how to ensure the solution adapts to market changes. We developed a recommendation system that ranks carriers based on their historical platform data using personalized k-nearest neighbour models and a custom similarity function. Tested in a controlled experiment, this system resulted in a 2.8% increase in the average matching rate, with improvements up to 3.4% in established regions and a peak of 6.2% in the Netherlands. However, the one-size-fits-all approach proved insufficient due to varying carrier capacities. To address the issue of deciding if and when to recommend, we introduced the Shipment Pulse Monitor (SPM), a decision tree-based predictive model that triggers recommendations when a shipment will likely be cancelled. To represent the trade-off between sending out recommendations to maximize matching probability and avoiding spam, we defined custom metrics for our problem. Our findings suggest that future work should focus on tailoring recommendations to specific carrier groups and combining hard cut-off points in data to allow a narrower focus for predictive models. Our work provides a comprehensive framework for improving matching rates on the UTURN platform with adaptive, targeted solutions.