In forensic investigations, an increasing amount of evidence is retrieved from digital devices. This evidence is often extracted from devices using digital forensic platforms. The platforms are able to extract digital traces from several types of files, originating from different applications, such as email applications from laptops or chat applications from smartphones. Developing support for a new file format and adding it to a digital forensic platform is time-consuming and difficult. An alternative approach is to integrate an existing digital forensic platform into another platform to extend its capabilities. This is, however, not trivial because firstly, it is difficult to choose a platform to integrate. This difficulty stems firstly from the fact that there exists no overview of open-source digital forensic platforms that can be used to compare advantages and disadvantages of the platforms. The first aim of this thesis is therefore to provide an overview of available open-source digital forensic platforms. This overview is created by means of online research.
The second reason that it is difficult to choose a platform to integrate is that it is difficult to assess whether the output of these platforms is accessible and structured such that it can easily be integrated. The second aim of this thesis is therefore to determine what the best method is to quantify the uniformity of the output of digital forensic platforms and to develop a proof of concept implementing this method. Developing the uniformity metric is done by first dividing the concept of uniformity into six sub-forms of uniformity. Conceptual methods to quantify each of those forms are provided, and we present a concrete implementation of a proof-of-concept.
The results of this thesis imply that although development of digital forensic platforms is actively ongoing, developers miss out on improving the digital forensic field as a whole by not considering the interoperability of the platforms they develop.

Efficient and inclusive data reuse across research disciplines is based on high quality metadata that bridges the gap between data producers and consumers. This gap, referred to as the metadata gap, arises when the metadata provided by producers do not meet the needs of consumers. Through a comprehensive analysis of metadata supply and demand, this thesis identifies the motivations and barriers faced by producers in creating metadata, along with the challenges consumers face when reusing datasets. To address these issues, the thesis introduces context-bridging data conversations, a framework designed to make metadata creation a more collaborative and adaptive process. The proof-of-concept is built on four key mechanisms: involving consumers as co-creators, recognising and incorporating contextual metadata, leveraging real-time dialogue, and dynamically adapting metadata elicitation questions. Qualitative interviews were conducted to identify the factors that shape metadata practices, and AI-generated summaries were evaluated as a scalable tool to synthesise the insights of these conversations. The findings are applied to the data management plans of CropXR, an interdisciplinary research institute. This case study illustrates how the metadata gap analysis can identify specific areas for improvement in metadata practices and how the context-bridging data conversations framework can provide actionable recommendations to enhance data reusability. By analysing data reusability as a dynamic and context-dependent process, this thesis advances both practical methodologies and theoretical understanding of metadata management. These contributions offer actionable strategies to close the metadata gap, foster collaboration across scientific domains, and promote more efficient and inclusive research practices.

With the advent of large language models (LLMs), developing solutions for Natural Language Processing (NLP) tasks has become more approachable. However, these models are opaque, which presents several challenges, such as prompt engineering, quality assessment, and error analysis. Explainability methods can have several potential benefits, such as improving accuracy, increasing trust, and assessing quality. However, limited research exists on how explainability techniques can be applied to LLMs in practice, particularly using human-centred methodologies. Therefore, this study takes a user-centered approach, investigating the needs and challenges of the NLP data scientist and developing an explainability tool to address these needs. This approach is done by conducting a formative study to deepen our understanding of the user, combined with relevant literature. The observations from the formative study were used to develop a tool tailored to the user’s specific needs. This development was done by creating requirements and a design based on the findings of the formative study, followed by a proof of concept implementation. User satisfaction was assessed through practical interviews with a fairness dataset, providing insights into the usefulness and usability of the explanation techniques and the tool. The tool implements three explanation techniques: uncertainty, token-level feature attribution, and contrastive explanations. These can be viewed using a web application separated from the Python development environment, making it easy to interact with. Other key features are that it can be easily integrated into the user’s existing workflow, is usable in practice and can be presented to different stakeholders within the project. The evaluation concluded that the tool fits the workflow and does indeed help the NLP data scientist to understand the model. However, the evaluation also showed that the explainability techniques did not provide the necessary insights to achieve the user’s goal, mainly to improve the model’s accuracy and make the error analysis actionable. More research should be done to see which other explainability techniques could provide insights that would lead to objectively better performance of these models. Finally, more explainability techniques should be developed that do not focus on debugging the model but rather on revealing its behaviour and thus providing a better understanding of how to improve it.

Large language models have achieved breakthroughs in many natural language processing tasks. One of their main appeals is the ability to tackle problems that lack sufficient training data to create a dedicated solution. Manga translation is one such task, a still budding and underdeveloped field, that at the same time deals with a problem even more complex than standard translation. One of its biggest challenges is successfully incorporating the visual modality in translation, to resolve ambiguities. Recently, a new type of LLMs -- multimodal LLMs -- has emerged and showed potential in understanding visual symbolism in narrative pieces like memes and comics. In this work, we investigate whether these models could be useful in the field of manga translation, ultimately helping manga authors and publishers make their works available to wider audiences. Specifically, we evaluate a number of methods based on GPT-4-Turbo for text only translation, image informed translation and volume-level translation. We perform both automatic and human evaluation. Moreover, we contribute new evaluation data -- the first parallel Japanese-Polish manga translation benchmark. Our findings show that our proposed methods are able to achieve state of the art results for English, and set a new standard for Polish. We conclude that while this is not a sufficient replacement for a professional human translator, it could help speed up the translation process or be used as a learning aid.

In recent years, the rapid advancements in big data, machine learning, and artificial intelligence have led to a corresponding rise in privacy concerns. One of the solutions to address these concerns is federated learning. In this thesis, we will look at the setting of vertical federated learning based on tree models. We have built a system that can do both entity resolution through private set intersection (PSI) and vertical federated learning (VFL). In this system, we have implemented an optimisation to pre-sort the data per feature before the start of VFL. We have also created a privacy framework, where we define four levels of privacy. This optimisation did not affect the privacy level of the system. In our results, we have seen that pre-sorting the data lowers the overall training time. How much depends on the number of entities and features of the passive party. We observe from our results that we estimate the speed-up to be 0.3654 seconds per feature and 0.2093 seconds per 1000 entities.

Music annotation and transcription of music sheets are traditionally performed by experts. Although these processes result in high quality data, the scope of each effort is relatively narrow resulting in highly specialised and specific datasets of annotated music compositions, which leads to a fragmentation in the design efforts for automated tools. In music traditions such as classical music, the shortcomings of current digitization work- flows become even clearer: due to the vast corpus and varying stylistic intricacies, experts tend to have specific knowledge and take up projects that concern very specific periods or composers, limiting our reach in regards to conserving classical music information as a whole.

On the other hand, crowdsourcing has been successfully utilized in other domains for annotating different modalities (text, image, video, audio), despite the unreliable pool of expertise on online platforms. Commercially successful projects have utilized the crowd, which provided annotations of adequate quality. These annotations were later used to fuel machine learning methods that rely on bulks of annotated data to perform automatic classification, regression, and detection tasks. However, due to the complexity of music as an artifact, there are still only a few examples where the crowd was integrated into any form outside of subjective annotation tasks (e.g., indicate the mood of the excerpt).

In this thesis, we tackle this research gap of integrating the crowd in the annotation processes of music compositions. We surveyed current practices on Optical Music Recognition to identify parts where the crowd could assist, alongside proposing hybrid annotation workflows for music compositions. We studied the capabilities of online participants with unknown musical expertise, quantified their musical abilities, and related them to their performance in music annotation tasks. With our goal being to identify ways to expand the preservation efforts for classical music through the assistance of the general public, we investigated potential online sources of music information and prospective participants outside the currently available crowdsourcing platforms. To that end, we studied how composers’ popularity manifests on community-driven platforms through the interactions of music enthusiasts online. We also conducted interviews and focus group discussions with experts and semi-experts, to understand their quality requirements on semantically-rich digital music scores, and identified transcription patterns that could inform our task designs. We finally delivered our system architecture, which combines computer vision and algorithmic scheduling, with microtasks designed to be performed by human annotators in parallel.

Our findings show that with the right methods to quantify the musical competence of a person, paired with careful design of the annotation tasks and interfaces, we can successfully integrate the crowd in the music annotation processes, to generate meaningful and useful information regarding classical music compositions and beyond. This thesis enables future research by showcasing the versatility of the crowd and providing task design methods to accommodate their lack of formal training in the field. It also provides experimental methods to reliably identify how different music composition elements affect the crowd’s performance, alongside proposing user interface elements that can mediate the complexity of the artifacts. Practices such as those presented in this thesis can lead to scaling up our digitization efforts, generating accurate and useful annotations through the crowd, even in such domain-specific and knowledge-intensive topics as classical music compositions.

Schema matching is a critical data integration process, which aims at capturing relevance between elements of different datasets; when datasets are tabular, it translates to the process of discovering related columns among them. Accurately discovering column matches is integral for several applications, such as entity resolution, data cleaning and data augmentation. While there exists a multitude of schema matching methods in the literature, we identify three major issues: i) there is no comprehensive study of comparing them in terms of effectiveness and efficiency, due to not available implementations and lack of evaluation datasets, ii) existing methods might be impractical and even inapplicable in certain modern settings, and iii) the heterogeneity and complexity of data can impede capturing relevance among columns for existing methods, as certain assumptions might not be holding for the entirety of underlying datasets. In this thesis, we tackle these issues by reviewing existing schema matching techniques and proposing novel methods capable to address challenges imposed by modern settings.
Starting with Chapter 2, we present an extensive comparison study on existing schema matching methods, by introducing Valentine. Specifically, Valentine constitutes an open-source experimental suite, which encompasses several state-of-the-art schema matching solutions. To guide the evaluation process towards modern applications, we extract four relatedness scenarios from the dataset discovery literature. To tackle the lack of existing datasets with ground truth, we devise a principled fabrication process. Our findings lead to insights that can help to improve future research on the field of schema matching, while they affect the design choices we make for novel methods we present in the following chapters.
Next, in Chapter 3, we turn our focus on applying schema matching among datasets stored in different data silos, which cannot be collocated and each contains information about column matches. Towards this direction, we introduce SiMa, a matching method that leverages existing matches in each silo, to build a column match prediction model, powered by the employment of a Graph Neural Network (GNN). To do so, SiMa transforms columns and matches among them in each silo to a graph, while it performs targeted negative edge sampling and incremental training to enhance the learning process. In our experimental evaluation, we show the benefits of using SiMa over state-of-the-art techniques, both in terms of effectiveness and efficiency.
Finally, Chapter 4 discusses the problem of discovering join relationships among datasets in a repository. To ameliorate the shortcomings of previous methods, we propose OmniMatch, a self-supervised method that can effectively capture both equi- and fuzzy-joins among tabular data. At the core of the method is the exploitation of a comprehensive set of similarity signals among columns, which are then transformed into a similarity graph. This graph, in conjunction with automatically generated positive and negative column match examples, enable the employment of a Relational Graph Convolution Network (RGCN) towards training a generalizable join prediction model. We compare the effectiveness of OmniMatch with several other state-of-the-art matching and column representation methods, while we verify the usefulness of utilizing a wide-spectrum of similarity signals to capture joins.
We conclude the thesis by reviewing our main findings, reflecting on our contributions and discussing potential limitations of the methods and approaches presented. Moreover, based on the insights we gain from surveying and developing novel matching methods, we discuss challenges and future directions in the field.

Despite the low adoption rates of artificial intelligence (AI) in respiratory medicine, its potential to improve patient outcomes is substantial. To facilitate the integration of AI systems into the clinical setting, it is essential to prioritise the development of explainable AI (XAI) solutions that improve the understanding of the AI predictions. These XAI solutions empower clinicians to collaborate effectively with AI systems, thereby enhancing the overall outcomes for patients in respiratory medicine. Unfortunately, the lack of user-centric studies in this domain has made it challenging to identify the specific aspects of explainability that are most effective in improving the adoption of AI in the real-world environment. To address this gap, we conducted a mixed-methods study of clinicians in respiratory medicine to identify the most relevant and crucial aspects of XAI solutions. Our study focused on understanding how XAI can be effectively translated into clinical practice by leveraging the expertise of doctors in the field. Because of the lack of knowledge about XAI concepts among pulmonologists a different approach is taken to regular user-centric XAI research and no direct examples of state-of-the-art XAI solutions are used. Rather the expertise of doctors is used to make them implicitly identify their needs and intents. Our findings reveal that the successful adoption of XAI solutions in respiratory medicine requires tailored solutions that address communication barriers, promote patient-centric care, and overcome AI adoption challenges. The study highlights the significance of task-specific visualisations, comprehensive explanations, preferred granularity, and the ability to mimic human judgement in successful XAI solutions. Trust and collaboration between clinicians and AI systems are essential for effective adoption, wherein AI is perceived as a colleague rather than a replacement. This ensures that clinicians can easily understand and work with the model predictions, ultimately leading to improved patient outcomes. By aligning XAI design with the needs and intents of pulmonologists, we established the importance of Co-designing solutions with domain experts and embedding XAI within clinical workflows emerged as key strategies. Our research underscores the imperative of transparency, extended validation, and continuous alignment of AI technologies with medical values. By following these principles, XAI solutions can be developed to enhance the diagnosis and treatment of respiratory illnesses, ultimately improving patient outcomes in respiratory medicine.

Colorectal cancer is a widespread disease that significantly impacts the health of individuals worldwide. Understanding the needs and concerns of those affected by this disease is crucial for improving patient outcomes and enhancing the quality of care. Patient web forums have emerged as valuable platforms for individuals to openly share their experiences and thoughts related to colorectal cancer, providing unique insights into the social, physical and emotional aspects of their patient journey. These forums offer a more comprehensive and authentic portrayal of patient experiences compared to traditional patient data collection methods, such as questionnaires and interviews, which may not capture the full scope of patients experiences in the colorectal cancer carepath.

However, analyzing the vast amount of unstructured data within these patient web forums presents a significant challenge. Traditional manual analysis by human experts is time-consuming, labor-intensive, and limited in scalability, making it impractical to analyze the sheer volume of patient-generated content. This is where the application of natural language processing (NLP) techniques becomes crucial. NLP enables the automated processing and analysis of textual data, allowing for efficient extraction and interpretation of the large amounts of patient forum posts.

Nevertheless, relying solely on machine intelligence, such as topic modeling and natural language generation, for interpreting patient forum data carries inherent risks, including the potential for disseminating misleading information. While these machine-driven techniques offer efficient and scalable ways to analyze and generate insights from the large amount of diverse and unstructured patient forums, they may lack the necessary contextual understanding and domain expertise to ensure the accuracy, relevance, and ethical implications for interpreting colorectal cancer patient experiences.

To close this gap between human experts and machine intelligence, this thesis explores the potential of hybrid intelligence (HI) for topic interpretation in colorectal cancer research. The main research question is: ``How can topic modeling, GPT-3.5 language generation and human expertise be combined to explore the interpretation of patient web forums in colorectal cancer (CRC) research?"

To address the research question, three human studies were conducted. The first study employed NMF topic modeling to compare topic interpretations created independently by medical workers and GPT-3.5. This comparative analysis discovered unique observations that differentiate human-written and AI-generated interpretations on online patient stories. In the second study, it was investigated how medical researchers collaborate with GPT-3.5 to develop hybrid interpretations on patient experience topics generated by the BERTopic model. A Flask web application served as the interactive platform for combining their knowledge with the AI model. Finally, the third study made professional human evaluators assess the topic relevance of the interpretations generated by medical researchers and GPT-3.5 to determine whether the combination of GPT-3.5 and human expertise leads to improved topic interpretations compared to individual interpretations.

The proposed solution to the research problem is to explore a hybrid workflow that compares, combines and validates GPT-3.5 language generation and human expertise, aiming for enhanced interpretations of topics extracted from colorectal cancer patient forums. The three studies provide opportunities for researchers and medical professionals to integrate machine intelligence from topic models and GPT-3.5 in their field of work. The hybrid workflow has conclusively demonstrated that human experts were successfully able to compare and enhance the relevance of human and GPT-3.5 interpretations of colorectal cancer patient experience topics. This allowed human experts to efficiently reach a more comprehensive understanding of patient forum data, which is essential for improving patient health in colorectal cancer research.

The goal of this research is to model and understand the effects of tourism demand on air quality by performing data integration on multi-source data. This research is aimed at researchers and practitioners aiming to perform multidisciplinary research in the fields of data science and geoscience, presenting the methods and challenges that arise when performing such an analysis. A data processing pipeline explains the research from a data integration perspective involving the data retrieval and pre-processing tasks. This enables the construction of datasets for machine learning modelling and prediction of air pollutant levels based on tourism data. The study area of this research is Italy which is chosen based on its significant tourism industry and wide availability of data about tourism development. For this study, in situ air quality data sampled using Google Earth Engine (GEE) around accommodation, transportation and tourism attraction locations is modelled with tourist arrival numbers, nights spent and average length of stay. Long short-term memory (LSTM) multivariate time series modelling is performed afterwards to understand predictability of air quality on a national and regional level. To this end, this research looks into three different stages of the modelling process of tourism with air quality which are: (i) retrieving accommodation, transportation and tourism attraction locations using the RDF model, (ii) identifying which pollutants are correlated and Granger-caused by the different tourism demand features using sampled satellite air quality data of the identified tourism locations, (iii) understanding performance characteristics of LSTM time series models by training on tourism demand and air quality data. Correlation analysis indicates the potential to model the relation between tourism demand indicators and PM2.5 in overall cleaner regions in terms of this pollutant. In these regions, Granger-causality testing suggests a higher chance of predictability of PM2.5 time series using tourism demand data from the previous month. Training an LSTM model using the information of this lagged relationship suggests that regions with overall high PM2.5 levels are challenging to model showing high RMSE scores. Training an LSTM model for these regions also required more epochs compared to overall cleaner regions to model the effects of tourism demand on air quality.

Gene similarity has been an area of great interest in numerous fields for decades, as it can provide insights into the evolutionary relationships among different species. This knowledge is particularly useful for advancing biotechnologies, discovering new drugs and treatments for various issues and improving the characteristics in breed crops or animals. DNA sequencing enables gene annotation, which facilitates the identification of similarities between genes. Similar genes from different species are interesting candidates for studying gene functional similarity. Gene ontology (GO) provides a standardized vocabulary for gene annotation, which is considered to be the ground truth when describing their properties. Nevertheless, an interesting source for gathering additional information about genes can be the plethora of biomedical articles accessible online. These are the pillars on which is foundered the incentive of this paper - an endeavor to investigate how using graph theory could benefit scientists in transferring knowledge about gene functionalities between different plants. We present an overview of the methodology and the design of the system we used in order to convey to what extent using subgraphs similarity based on annotated data proves to yield results similar to those already established as ground truth for the model plant Arabidopsis Thaliana and its counterpart, Solanum Lycopersicum, as well as our conclusions and discussions regarding the quality of the datasets used throughout this research.

There are an estimated 253 million blind and visually impaired people in the world. To grant them access to text publications that contain images, experts are employed to write image descriptions. There is both a societal and a legislative pressure to supply image descriptions to all new and archived publications within a number of years, yet the number of available experts is limited. The image description task for images in textual context is complex, since a fitting description combines salient elements from both the image as well as the context into a description, which also differs depending on the publisher and published medium.

Because of this complex nature, current automated systems are unable to reliably produce desirable results. Instead, this thesis focuses on developing software to assist the experts in their general image description task in order to improve their efficiency. Specifically, we use existing, commercially available automated tools to generate alternative representations of the data. To analyse the system, we develop a user interface to present all of the available data and design an experiment with a small group of experts to investigate the system's applicability and perceived usefulness. We find that such a system has great potential to assist the experts, but that it might be desirable to focus on a solution aimed towards a smaller subset of publications, so that domain-specific information sources can be exploited to improve the information quality.

Over time Linked Data collections are continuously subject to change because of numerous reasons. Users could insert new observations, or they could rectify erroneous statements in these knowledge graphs. In order not to lose historically import information, this trend of evolving Linked Data collections increases the need to version these collections. Furthermore, retrieving prior versions and their in-between changes could provide Linked Data users relevant information. However, for some changes to these collections we should record both their transaction time and their valid time. To address these two problems of versioning Linked Data collections and having bi-temporal changes, we introduce the Bi-Temporal Versioning Approach for Knowledge graphs (Bi-VAKs): a prototypical bi-temporal change-based Version Control System for an arbitrary RDF dataset. Bi-VAKs registers both the transaction time and the valid time of a set of modified quads, and therefore, it allows for coupled historical and retrospective SPARQL queries. In addition, in order to enhance collaboration between its users Bi-VAKs also keeps track on provenance data; it supports diverged states; and provide a standard data access interface. However, since the standard RDF data model is atemporal, defining such a set of modifications (update) in RDF poses difficult challenges. Firstly, to indicate this transaction time and the valid time of an update Bi-VAKs divides a revision or version into a transaction revision and a valid revision. And hence it directly separates the metadata from the actual data. Secondly, to denote and retrieve the modified triples/quads within a update Bi-VAKs uses RDF-star and SPARQL-star. In order to connect these revisions we develop three reference strategies: the explicit, the implicit, and the combined reference strategy. These strategies let a transaction revision either refer explicitly to its corresponding valid revision(s) or implicitly by the same revision number and branch index. Based on these strategies we initiate some different approaches to query the updates from the revision-store. And, we propose some different methods to construct a (prior) version. However, to evaluate these different design decision we cannot use the existing uni-temporal benchmarks for our bi-temporal versioning approach. Therefore, we expand the BEAR benchmark to a bi-temporal benchmark (Bi-BEAR). By means of this benchmark we demonstrate that all three reference strategies have about the same storage size. We notice that the usage of a snapshot and retrieval of all updates worsen the version materialisation (VM), delta materialisation (DM), and version (VQ) query performance. In addition, the VM query look up time considerably decreases if only the matching updates are queried. And modified updates, branches, and more updates in the revision-store slightly lower the VM, DM, and VQ query performance. In addition, for the implicit and combined reference strategy the query time is rather the same, and sometimes even better if we sort the updates instead of aggregating them directly. Overall, the implicit reference strategy is performing best, and is quickly followed by the combined reference strategy.

Public technology has been shown to have a strong dependence on physical touch, which increases the transmission of diseases. Gesture recognition helps to reduce this transmission, as the dependence on physical touch is removed. Furthermore, the use of visible light for gesture recognition would reduce the power consumption of public technology, as less power would have to be supplied for a light source. In this paper, a dataset for gesture recognition using ambient light is presented, alongside the design process and challenges faced. The gestures were collected using three photodiodes in order for a machine learning algorithm to identify the patterns made by the shadows cast when the gestures are performed. The dataset consists of 10 gestures performed by a total of 50 people 5 times on each hand. This was collected under 5 different light intensity ranges. This dataset was then passed through a machine learning model to be trained and tested, resulting in a 86.8% (3.s.f) validation accuracy. There are many factors related to the light source that caused the accuracy of the algorithm not to be as high as expected, however, the highest accuracy was found in environments with light intensities of 100-1000 lux; a well-lit indoor room.

Dataset discovery techniques originally required datasets to have the same domain which made them unsuitable to be used on a larger scale. To avoid this requirement, newer techniques use additional information, aside from the datasets being processed, to better understand the data. They might rely on a knowledge base that describes the meaning of data, or a lexical database that defines meaningful relations between the words contained within the data. The main problem with this approach is that these types of additional information have poor coverage of the data being analyzed.

I propose to use a type of information that I call dataset domain terms. These are terms, or data values, that represent the domain of a dataset. I provide a technique that can derive these dataset domain terms automatically from existing datasets which means they are easily available. The problem of poor coverage can also be mitigated by only discovering datasets for the domain represented by these dataset domain terms. I provide a dataset discovery technique that takes this approach with these dataset domain terms.

Through an evaluation, I show that these dataset domain terms are sufficiently representative of the domain to be used for dataset discovery. The accuracy of the dataset discovery technique is also shown to be comparable to state-of-the-art dataset discovery techniques, though its precision is lacking.

This makes it highly suitable to filter datasets before other dataset discovery techniques can be performed on them. The data from these filtered datasets should also have a limited range of domains. So subsequent dataset discovery techniques should be less affected by the poor coverage of the additional information they use to understand the data. This allows dataset discovery to be performed on a larger scale.

Binary Neural Networks (BNNs) are compact and efficient by using binary weights instead of real-valued weights. Current BNNs use latent real-valued weights during training, where several training hyper-parameters are inherited from real-valued networks. The interpretation of several of these hyperparameters is based on the magnitude of the real-valued weights. For BNNs, however, the magnitude of binary weights is not meaningful, and thus it is unclear what these hyperparameters actually do. One example is weight-decay, which aims to keep the magnitude of real-valued weights small. Other examples are latent weight initialization, the learning rate, and learning rate decay, which influence the magnitude of the real-valued weights. The magnitude is interpretable for real-valued weights, but loses its meaning for binary weights.
In this paper we offer a new interpretation of these magnitude-based hyperparameters based on higher-order gradient filtering during network optimization. Our analysis makes it possible to understand how magnitude-based hyperparameters influence the training of binary networks which allows for new optimization filters specifically designed for binary neural networks that are independent of their real-valued interpretation. Moreover, our improved understanding reduces the number of hyperparameters, which in turn eases the hyperparameter tuning effort which may lead to better hyperparameter values for improved accuracy.

Investigating changes in forest cover has been an area of intense research for decades. From manual surveys to remote sensing we have come a long way in mapping the world around us. Machine learning and its' younger sibling, deep learning, have emerged as highly useful tools on this journey. There are a wide variety of algorithms that take a stab at analysing vegetation coverage, but not all of them are easily accessible to the community at large. This was the underlying incentive for the paper you are about to read - reproducing and replicating a novel deep-learning algorithm that generates maps of forest cover in a relatively under-explored part of the world. This study lays out the methodology devised by the original authors as well as our attempts to replicate their work on a different set of data. We present the results we were able to obtain, the hurdles we encountered along the way, and a set of guidelines we feel would be helpful for future researchers.

Electrocardiography is the craft of producing electrocardiograms. These graphs give physicians insight into the potential pathology of the heart. In order to come to a diagnosis, physicians use electrocardiograms in combination with follow-up physical examinations. There has been extensive research into automated methods that can differentiate healthy individuals from pathological individuals when given only an electrocardiogram. Some of these methods make use of neural networks that do feature extraction and classification in an end-to-end fashion. Related research performs feature engineering on ECG data followed by clustering. The evaluation shows that the resulting clusters coincide with heart pathology annotations. This master thesis describes the search for a machine learning pipeline, where a combination of an input representation, autoencoder, and clustering algorithm produces clusters coinciding with heart pathology without being biased by either heart pathology annotations or feature engineering that is already known to be predictive of heart pathology. Although the preexisting methods yield state-of-the-art accuracies, they do not allow us to learn about the structure and patterns in the data and the supervised methods are very costly to create. There has been no research into feature extraction by autoencoder followed by unsupervised clustering. Doing this could expose patterns in the data that could lead to improved diagnostics of heart pathology, and to automatic methods that are cheaper to create by turning the problem of diagnosis into an unsupervised or semi-unsupervised problem. In order to find the combination of input representation, autoencoders, and clustering that is best suited for predicting heart pathology from ECG signals, experiments are done that give insight into how much the resulting clusters coincide with heart pathology. The experiments work by first feeding some representation of one-second ECG objects into the autoencoder, after which the resulting low-dimensional representations are fed into a clustering algorithm. The clustering algorithm gives every low-dimensional ECG object a cluster label. The cluster labels are mapped to heart pathology labels by making use of existing heart pathology labels. This heart pathology label can now be interpreted as a class prediction in the context of classification. This setup is created to quantitatively answer to what degree the resulting clusters coincide with heart pathology. In the end, the concatenated image plot representation objects fed into a convolutional autoencoder followed by SOM clustering are compared to existing research. The classification accuracy achieved by the autoencoder pipeline formed in this research is 0.76 ± 0.01. This means that the clusters formed by the existing research coincide much more with heart pathology labels than the clusters from this research. A qualitative visualization from the low dimensional representations after the autoencoder, however, shows that the setup from this research is better at identifying patient IDs than heart pathology. This means that the features extracted by the autoencoder are salient for identifying persons, but not for identifying heart pathology.

Machine Learning models are begin increasingly used within the industry such as by financial institutions, governments and commercial companies. In the past few years, there have been several incidents where these ML models show discriminatory behavior towards particular groups of people, leading to unfair decisions that can have negative impacts on the lives of these people. Therefore, eliminating bias and ensuring fairness within these models is crucial to the societal expectations these institutions would like to meet. While there are tools and research towards technical mitigation methods for bias and unfairness. There is a lack of focus on the process of implementing these tools and methods within the industry, to develop ML models with the consideration of fairness at an early stage. In particular, there is a lack of specification on what fairness goals and objectives the ML model should accomplish. Without having this clarity, industry stakeholders can apply the tools and algorithmic unfairness mitigation methods but if the fairness requirements are not defined, then a) one is still not sure whether the ML model is solving the correct fairness goals and b) one is still not sure what the trade-offs and feasibility within different fairness goals, and available resources may look like. To address this, we design a Multi-Level Fairness Framework (digital workflow) that aims towards supporting stakeholders within the industry to perform Requirements Engineering, specifically elicitation and modeling, for fairness in a Machine Learning Model.

Database entity type recognition is the practice of recognizing conceptual entity types for which given data sets contain data. In big data or data lake settings, it is not always known which conceptual entity types are represented in each data set, making it difficult to extract value from the data. Depending on the logical schemas, each conceptual entity type can also be represented in the data instances in multiple different ways. This phenomenon, called semantic heterogeneity, poses a challenge when attempting database entity type recognition. Narrowing down the problem space to a specific organization makes it easier to cope with such problems. Organizations know which entity types are used in the organization and require only those to be recognized. And while there is heterogeneity in representation, there is likely a common set of rules each logical schema adheres to which can be exploited to recognize semantic heterogeneity. Furthermore, experts at an organization can provide example data instances for each conceptual entity type of interest, which provide ground truth for the proposed database entity type recognition solution. The proposed solution makes data profiles of the example data instances, and then attempts to recognize entity types in previously unseen data instances using a rule-based approach. Rules are used to maximize the ease of explainability of results, as is often desired at a bank, and can easily be added to or removed from the solution to maximize adaptability. Experiments using the proposed solution show promising results, with up to 90 percent of entity types correctly recognized over a total of 170,000 entities.