Interpretable models are essential in many machine learning applications, particularly in domains where transparency and trust are critical. Decision trees are a popular interpretable model, but their structure often leads to repeated identical subtrees and data fragmentation, which can result in large, overfit models with poor generalization. Decision diagrams could offer a more compact and less fragmented alternative by allowing sharing parts of the diagram, but constructing decision diagrams is computationally challenging and has seen limited practical adoption. This thesis introduces a novel local search-based algorithm for learning binary decision diagrams for classification. Our approach finds a middle ground between early greedy methods and exact optimization techniques, enabling scalable construction of compact and accurate diagrams. We define a local search approach with several move operators and explore multiple metaheuristics, identifying hill climbing with information gain-based initialization as the most effective strategy. We refer to this method as Decision Diagram Local Search (DDLS). We evaluate DDLS on 57 real-world datasets from the UCI repository and 480 synthetic datasets generated from known diagrams. Our method achieves competitive or superior accuracy compared to state-of-the-art decision tree and diagram methods on real-world datasets, while producing small models with lower fragmentation. Though challenges remain, especially on complex synthetic datasets, our results suggest that DDLS, and decision diagrams in general, hold significant untapped potential as interpretable classifiers.

Dimensionality reduction (DR) algorithms have been proven useful in various tasks when it comes to exploring high-dimensional data. Being one of the most used DR techniques, t-SNE is often valued for its ability to map nonlinear manifolds and preserving local structures. However, t-SNE produces anonymous axes which lack interpretability. To enhance the interpretability of t-SNE, this work proposes DimenFix-t-SNE, an extension of the existing meta-strategy DimenFix. DimenFix-t-SNE implements DimenFix using t-SNE, explicitly preserving the selected input feature in one of t-SNE's output dimensions. We generalize upon DimenFix by introducing individual ranges for every point and allowing more user control on the "pushing" process. For fixing nominal features in particular, we develop a new pushing mode Rescale, as well as an ordering method for the feature along the fixed axis. The extended method also supports traceable switching between features on the fixed axis, adding to the interactive component. Finally, we implement and test the algorithm as a plugin for ManiVault, providing an interactive and explainable visual analytics tool for high-dimensional data.

Introduction: In critically ill paediatric patients, sleep is essential for recovery and development, yet sleep disturbances are common in the paediatric intensive care unit (PICU), which highlight the need to integrate sleep monitoring into clinical practice. While automated sleep stage classification using machine learning (ML) on single-channel electroencephalography (EEG) data has shown promise in mostly healthy adult populations, its application to critically ill children is challenged by age-specific sleep architecture, medication effects, pathological conditions, and artefacts. This study evaluates whether deep learning (DL) feature extraction and dynamic models can improve sleep staging performance in this population and explores the use of an unsupervised ML model to gain deeper insight into the complex sleep structures in this population.

Methods: This study utilised EEG recordings from three datasets—healthy adults, non-critically ill children, and critically ill children—to train and evaluate supervised and unsupervised sleep stage classification models. As supervised models, a convolutional neural network (CNN) was used for feature extraction, followed by dynamic models including a long short-term memory (LSTM) network and a hidden Markov model (HMM) to account for temporal dependencies in sleep. Additionally, an unsupervised HMM was applied to explore underlying structures in the sleep EEG data without predefined labels.

Results: Supervised models achieved good performance in healthy adults and non-critically ill children, with maximum accuracies of 90.2\% and 77.4\%, respectively, for three-state classification. The added value of dynamic models over the CNN alone varied per dataset and model type and was not consistent. In critically ill children, classification performance was low, with a maximum accuracy of 61.4\%, and notably low macro-F1 and Cohen’s kappa scores (45.9\% and 26.5\%, respectively). The unsupervised HMM revealed that identifying distinct and stable clusters over time was challenging in all datasets. For critically ill children, the model often failed to identify multiple distinct clusters within individual patients, and substantial variability in cluster assignments was observed across patients.

Discussion: This study demonstrates that DL-based feature extraction and dynamic modelling using single-channel EEG can achieve strong sleep staging performance in healthy adults and non-critically ill children, highlighting the potential for (semi-)automated scoring tools in more stable populations. In contrast, performance in critically ill children was notably lower, likely due to factors such as high variability in sleep architecture, signal artefacts, limited data quality, and the uncertain reliability of manually assigned labels. These results suggest that conventional sleep stages do not generalise well to this population, and a purely data-driven, unsupervised approach does not offer a viable alternative. Overall, the findings emphasise the need for a larger dataset of critically ill children, further evaluation of relevant patient and data characteristics, the inclusion of alternative signals such as electrocardiography, and greater focus on model interpretability.

Machine learning (ML) is increasingly integrated across diverse academic disciplines, necessitating effective teaching strategies tailored to varied student backgrounds. This study investigates the influence of prior mathematical knowledge on the learning outcomes of ML topics among Computer Science (CS) and Aerospace Engineering (AE) students. Employing a mixed-methods approach, the research involved initial mathematical assessments, interactive tutorials on key ML topics (Bayes Rule, Perceptrons, ML Pipelines), and subsequent evaluations of ML comprehension.

The results reveal significant differences in performance between the two groups. CS students, with their integrated programming and mathematical preparation, consistently outperformed AE students, who demonstrated variability despite their strong quantitative foundations. Probability and linear algebra emerged as key contributors to ML learning, showing stronger correlations with outcomes than calculus. Qualitative analysis highlighted the need for tailored instructional approaches: AE students preferred application-driven and interactive learning, while CS students valued structured and technically detailed resources.

These findings underscore the importance of interdisciplinary teaching strategies that bridge gaps in programming and mathematical competencies. The study’s insights have implications for designing inclusive ML curricula, emphasizing real-world applications, adaptive learning technologies, and frameworks to support diverse learner needs. Future research should explore broader ML topics, larger participant groups, and long-term skill retention to further enhance ML education across disciplines.

Machine learning (ML) has become a vital skill across various disciplines, driving innovation and transforming industries. This growing demand emphasizes the need for effective teaching methods tailored to students with diverse academic and technical backgrounds. Teaching ML to non-majors presents significant challenges, as many students lack foundational knowledge in mathematics and programming. This study explores how university instructors addressing these challenges in bachelor’s and master’s level courses, based on semi-structured interviews. The analysis uncovered several key strategies used by instructors. Many emphasized connecting ML concepts to real-world applications, making the subject more relatable. Visualization tools were commonly employed to simplify abstract concepts and improve comprehension. Hands-on activities, including live coding and interactive assignments, were highlighted as effective methods to engage students and bridge theory with practice. However, instructors faced challenges as well such as accommodating diverse student backgrounds, correcting misconceptions, and designing assignments that balanced accessibility and depth. Although tracking student progress was not the focus of this study, some insights were provided. Some instructors mentioned using formative assessments, such as quizzes and project-based evaluations, to measure understanding. These findings highlight the importance of adaptable teaching methods and inclusive learning experiences in making ML education more accessible and effective for non-majors, while also providing actionable advice to improve the educational process and course design.

As Machine Learning (ML) continues to shape advancements in academia and industry, ensuring effective ML education is essential. This study examines the retention of four core ML concepts- Principal Component Analysis, Gradient Descent, Bayes’ Theorem, and Hierarchical Clustering- two years after students completed a university-level ML course. Using a survey-based methodology, it explores how prior mathematical knowledge, perceived difficulty, and confidence influence long-term retention. Results reveal a significant positive correlation between Calculus knowledge and Gradient Descent retention, with weaker correlations for Linear Algebra with PCA and Probability with Bayes’ Theorem. Perceived difficulty and confidence also shape retention outcomes. The findings emphasize the need for targeted mathematical refreshers in ML courses to strengthen foundational knowledge and improve retention. This research provides actionable insights for curriculum design, aiming to bridge mathematical gaps, enhance learning outcomes, and sustain student engagement with advanced ML concepts.

With the growing need for machine learning knowledge for many different expertises and positions, comes a growing need for machine learning education for non-computer scientists. Teaching machine learning concepts to non-majors comes with the added challenge of dealing with different levels of prior mathematical knowledge. Existing research is inconclusive on the correlation between this prior knowledge and topic-specific machine learning knowledge gain. This paper evaluated this via an experiment conducted on Computer Science and Physics students without prior machine learning education. We find that there is no clear correlation between general math knowledge and knowledge gain. There is however a clear correlation of proficiency in probability and statistics, and algorithm heavy machine learning topics. The experiment also concluded that most students struggled most with these math-heavy topics, as well as understanding abstract systems such as perceptrons.

An accurate segmentation model for hip compo- nents could improve the diagnosis of Osteoarthritis, a prevalent age-related condition affecting joints. A significant challenge in developing effective and robust segmentation models are the domain differ- ences across various datasets. In this study, we in- vestigate the impact of different data augmentation and preprocessing techniques on the generalizabil- ity of femur segmentation models across datasets. Using two labeled datasets, we evaluate the perfor- mance of a U-Net segmentation model, focusing on the effectiveness of augmentations like image flip- ping, random rotations, blur, contrast, and bright- ness adjustments. Our findings reveal that certain augmentations, particularly random rotations of up to 15 degrees, vertical image flipping and light blurring, significantly improve the model’s gener- alization to another data set, reducing boundary er- rors and enhancing segmentation accuracy. These results underscore the potential of targeted data augmentations in developing robust, generalizable models for hip joint component segmentation.

With the fast integration of Machine Learning(ML) into several industries, the motivation to develop effective pedagogical strategies for teaching this complex and evolving field has become critical. Machine Learning, once mainly a topic in Computer Science Bachelor programs, is now widely integrated into various majors and introduced at earlier educational stages, including high school and secondary school. However, the reduced research focus on ML pedagogy results in a lack of standard teaching methods compared to other science-related subjects, which have established norms for topic introduction, teaching tools, and assessment methods. With inspiration from other fields, this research aims to look into using interactive visualizations in teaching ML topics, more specifically in teaching Gradient Descent and Principal Component Analysis (PCA). The research includes the development and evaluation of Jupyter Notebooks for introducing these visualizations to students. The targeted student population is composed of Computer Science and Engineering Bachelor students who have not yet followed any Machine Learning courses but have the necessary background knowledge, namely calculus, linear algebra, and statistics knowledge. The evaluation of this teaching method measures the knowledge gained and the motivation of the students, compared to a static version of the materials. The results of the evaluation have shown a significant effect of interactive visualizations on knowledge gained related to PCA. The evaluation has not identified any difference in knowledge gain for gradient descent and learning motivation for both topics. With these results, we contribute to the body of evidence-based teaching methods in Machine Learning and identify further research needed to generalize the effect of interactive visualizations as a teaching method for teaching ML basic concepts.

Mechanical ventilation is a vital supportive measure for patients with acute respiratory distress syndrome (ARDS) in the intensive care unit. An important setting in the ventilator is the positive end-expiratory pressure (PEEP), which can reduce lung stress but may also cause harmful side effects. This research investigates the personalization of PEEP settings based on patient characteristics using three meta-learning algorithms (S-, T-, and X-learner) to estimate the conditional average treatment effect. Additionally, the hypothesis that the X-learner performs particularly well under a significant imbalance in patient numbers between treatment groups is tested. Results show that the X-learner slightly outperforms the S- and T-learners in terms of mean squared error under various unbalanced conditions in simulated data. However, the overall ability of these meta-learners to identify patients benefiting from high PEEP remains inconclusive. When using gradient boosted trees or random forest as base models, cumulative gain curves on MIMIC-IV data indicate potential overfitting. While the X-learner performs somewhat better on this data, the low area under the curve scores suggests a minimal distinction between high and low PEEP groups. External validation with data from a randomized control confirms that the models do not effectively distinguish between treatment groups. These findings suggest that further investigation with more complex models and real-world data is needed to validate the potential of meta-learning algorithms in personalizing PEEP settings for ARDS patients.

Osteoarthritis (OA) is a chronic musculoskeletal joint disease that leads to disability. Osteophytes are a hallmark of OA in the knee, characterized by the formation of bone spurs that contribute to joint pain and reduced mobility. This study explores the application of deep learning (DL) techniques for the automatic detection and grading of osteophytes on magnetic resonance (MR) images of the knee. Leveraging the DenseNet-121 and ResNet-50 DL architectures from the Medical Open Network for Artificial Intelligence (MONAI) framework and a dataset, containing 1782 double echo steady-state (DESS) MR images from the Osteoarthritis Initiative (OAI), the study aims to enhance diagnostic accuracy and efficiency in medical imaging analysis. The dataset was split 8:2 for training and validation purposes, respectively. Through a series of numerical experiments, the research evaluates binary classification, region of interest (ROI)-based detection, and multi-class classification models, demonstrating that DenseNet-121 generally outperforms ResNet-50. The five-fold cross-validated binary DenseNet-121 model achieved an area under the receiver operating characteristic curve (ROC AUC) score of 0.90 and a balanced accuracy of 0.82, with a 95% confidence interval (CI) of 0.81-0.83 trained on resampled whole knee images. Moreover, the cross-validated ROI detection models for the patella inferior, superior, and tibia lateral subregions achieved balanced accuracy scores of 0.89 (0.88-0.90 CI), 0.86 (0.85-0.87 CI), and 0.85 (0.84-0.86 CI), respectively. However, the multi-class DenseNet-121 model achieved lower performance, with a balanced accuracy of 0.73 (0.71-0.75 CI), indicating the complexity of multi-class classification in this context. Furthermore, this research did not include hyperparameter optimization, as many settings were kept at their default values, suggesting the possibility for improved results. The cross-validated models were evaluated on an external test set, obtained from the Erasmus Medical Centre, comprising FSPGR-FS images from a significantly younger patient cohort, with a notable class imbalance. The models’ performance on this dataset was significantly lower than their validation results, underscoring the limitations in generalizing to different age demographics and class distributions. External testing underscores the need for more robust models to maintain high performance across diverse datasets and clinical settings. Key contributions of this study include the use of weighted categorical cross-entropy (WCCE) loss functions and analysis of the knee’s subregions to improve detection accuracy. The findings establish a solid foundation for further research, suggesting future work should focus on advanced optimization techniques, mixed imaging sequences in the training dataset, and comparative studies with other established models within the computer vision sector.

In problems with missing data, the data are often considered to be missing at random. This assumption can not be checked from the data. We need to assess the sensitivity of study conclusions to violations of non-identifiable assumptions. This thesis performs Bayesian sensitivity analysis for a missing data model with life time outcomes and covariate information. The outcome distribution is modelled through a Cox model, with a beta process prior on the cumulative hazard function. We run experiments in a simulation study to test the performance of the model in scenarios with simulated data of several sample sizes. We show the validity of the model in the context of Bayesian sensitivity analysis, and propose extensions.

Learning curves in machine learning are graphical representations that depict the relationship between a model's performance and the amount of training data it has been exposed to. They play a fundamental role in obtaining the knowledge and skills across a range of domains. Although there are already quite some researches studying machine learning curves, explaining the importance and practical application of learning curves, we still know very little about the factors that influence the parameters of the learning curve. The aim of this research is to give a better understanding of different factors affecting the parameters of the learning curve. Specifically, we are interested in how the dimensionality of a dataset can influence the parameters of the learning curve. Since learning curves are useful and have several applications, such as estimation of the time required to complete production runs, we would like to know if the dimensionality has any effect on the shapes of learning curves. To conduct the research I applied principal component analysis (PCA) three times with different amount of information preserved to reduce number of dimensions on several datasets and analysed the changes in the parameters of the obtained learning curves. The research showed that potentially there might be some relation between dimensionality and shape of the curve, but only in cases of specific machine learning model. The amount of experiments conducted is not sufficient to make solid conclusions and it is advised to continue with proposed experimental setup, but train machine learning models on increased number of datasets.

Concept drift is an unforeseeable change in the underlying data distribution of streaming data, and because of such a change, deployed classifiers over that data show a drop in accuracy. Concept drift detectors are algorithms capable of detecting such a drift, and unsupervised ones detect drift without needing the data’s actual labels, which can be expensive to obtain. This work is concerned with the implementation and evaluation of two existing unsupervised concept drift detectors based on clustering, UCDD and MSSW, by evaluation on both synthetic and real-world data. Our biggest contribution is in making implementations publicly available. By evaluation, we also realise that UCDD detects drift earlier for simple numerical synthetic datasets, MSSW detects drift earlier for more complex synthetic datasets with categorical features, and none seems suitable for real-world datasets.

Various techniques have been studied to handle unexpected changes in data streams, a phenomenon called concept drift. When the incoming data is not labeled and the labels are also not obtainable with a reasonable effort, detecting these drifts becomes less trivial. This study evaluates how well two data distribution based label-independent drift detection methods, SyncStream and Statistical Change Detection for Multi-Dimensional Data, detect concept drift. This is done by implementing the algorithms and evaluating them side by side on both synthetic and real-world datasets. The metrics used for synthetic datasets are False Positive Rate and Latency; for real-world datasets, Accuracy is used instead of Latency. The experiments show that both drift detectors perform significantly worse on real-world than on synthetic data.

When deployed in production, machine learning models sometimes lose accuracy over time due to a change in the distribution of the incoming data, which results in the model not reflecting reality any longer. A concept drift is this loss of accuracy over time. Drift detectors are algorithms used to detect such drifts. Drift detectors are important as they allow us to detect when a classification model becomes inaccurate. Some possible uses of drift detectors can even go as far as detecting adversarial attacks on machine learning algorithms. The detectors discussed in this paper are Margin Density drift detectors. Their evaluation is made within an unsupervised context, where we assume no testing labels are available. In real world applications of machine learning models, this might often be the case, as finding labels is costly. Experiments in this paper have found that margin density detectors can be useful tools in detecting the first drift for synthetic data, even though parameter tuning must be done to achieve high accuracy for some datasets. In an unsupervised environment with more than one drift, the drift detectors are unreliable as was seen in experiments involving real world data. With this paper comes an implementation of margin density detectors.

Nowadays, software is an integral part of many companies. However, the codebase can grow large and complicated and is often insufficiently documented. To gain insight, tools have been made to infer state machines and process models from software logs. These tools produce different types of models such as automata and Petri nets. The main objective of this research is to determine which tool is the optimal choice for inferring a readable and correct model within reasonable time. Currently, Petri nets and automata are not compared to each other and not all key performance indicators are applicable to both model types. To compare these different concepts, suitable metrics must be identified.

For this work, 8 configurations of 4 programs will be compared. Finite State Machines (FSMs) will be inferred with FlexFringe (AIC), MINT and PRINS (using MINT internally). Petri nets will be mined with ProM using the Inductive Miner, Inductive Miner Infrequent - All Operators, Hybrid-ILP and the Directly Follows miner. The configurations will use 5-folds cross validation to infer models on 9 software logs. Negative traces were synthesised as they were not available. The quality of the models will be measured though inference time, complexity, F2-score, balanced accuracy, fitness and perplexity.

Some of the used metrics were adequate, but others were not suitable. Inference time, F2-score, balanced accuracy could be measured for both FSMs and Petri nets. The complexity was measured with the Petri net $eCFC$ metric and the Cyclomatic complexity CC. The eCFC does not properly express complexity on FSMs. Furthermore, Petri nets can model parallelism, which introduces extra complexity compared to an FSM. This was not adequately expressed by either of these metrics. To measure fitness, both token-based replay and alignment fitness were used. FSMs were converted to Petri nets. Token-based replay (TBR) fitness was not an expressive metric for the FSMs, as the concept of tokens did not carry over well. In addition to this, the external implementation of TBR fitness was flawed for the specific structure of the converted FSMs. Alignment-based fitness is the superior fitness metric as it does not rely on the notion of tokens, which the FSMs do not have. Unfortunately, the time and memory needed for alignment computations was too large for some models. Lastly, the perplexity FSM metric was successfully adapted for Petri nets. It expresses the difference in structure and could be tailored even further for the purposes of comparison by adjusting its parameters.

The results of the comparison showed that almost all configurations could complete inference in feasible time and memory. The time out was set at 4 hours and the available memory was \texttt{16GB}. The MINT and PRINS ran out of memory on one of the larger logs and timed out for one other set. Hybrid-ILP timed out for 2 sets. All other configurations completed inference for all data sets in under 40 seconds. PRINS and MINT boasted excellent performance across all correctness metrics, and were only outperformed by FlexFringe and the Directly Follows miner on perplexity. PRINS and MINT were most suitable for modelling traces of data sets with a low trace similarity and generalising to identify new traces. However, MINT models were a lot larger than those of any other configurations and PRINS models are generally many times larger than MINT models. So, if complexity of models is a big concern, FlexFringe and the Directly Follows miner offer the smaller models, at the cost of a small amount of performance for most sets. However, these two tools perform poorly on sets that are both incomplete and contain dissimilar traces. If time is of the essence, one should use FlexFringe, the Directly Follows miner of one of the Inductive Miners. The Petri net miners used, were not designed to introduce new behaviour in a model. Therefore, FlexFringe is preferable for modelling an incomplete log.

Flood simulations can give insight into the consequences of flood scenario's and can help to create hazard- and risk maps to support decision-making in flood risk management and in crisis management. 2D hydrodynamic simulations give accurate descriptions of the propagation of a flood and rely on advanced numerical methods to solve a set of physics-based mathematical equations. A drawback of these models is that they can be computationally expensive with run times in the order of hours or days depending on the time and spatial resolutions. In this study we explore the use of deep learning techniques in a surrogate model for 2D flood simulation. We propose and test a deep learning-based surrogate modelling framework that can be used to train a deep learning-based surrogate model. Once trained, the surrogate model can be used as a substitute for the hydrodynamic model with the advantage of being much more efficient in terms of run time and can be of great value in for example crisis situations. For training, a data set of expensive 2D hydrodynamic simulations was created using the SOBEK software program. Such simulations require a lot of input data, such as input parameter maps specifying the terrain over the computational grid and boundary conditions. To make training data-efficient, a sampling strategy was used for the input of the flood simulations. Three deep learning architectures were trained and tested. The first two architectures are feed-forward networks and the third architecture is of recurrent network type. These networks contain convolutional neural network (CNN) architectures with an encoder-decoder structure to make patch-level predictions of the flood characteristics in time. These patches contain a small section of the flood prone area and an encoder network is used to extracts coarse feature maps from this data that is then refined by a decoder network to create a prediction of the flood propagation. Using patches has the advantage of making a surrogate model able to create flood simulations over prone areas without restrictions on size or shape by tiling the output patches with flow predictions. Also it allows the surrogate to focus only on regions where the flood has reached and not on the regions where no water has arrived. It was found that with the recurrent architecture, the surrogate model was most capable of emulating the ground truth flood simulations in the test simulation. This trained network architecture was used in a case study where the surrogate was applied to create flood simulations in a small dike ring in the Netherlands. This shows that the surrogate modelling framework can be used to train a deep learning-based surrogate and, once trained, can be used to create flood simulations similar to hydrodynamic simulations. However, two main challenges were identified in using such data-driven deep learning-based surrogates. Firstly, keeping the predictions of the flood characteristics accurate enough to avoid large error propagation. Secondly, accurately generating large amounts of data from relatively little information present in the boundary condition and terrain.

Year after year, the amount of network intrusions and costs associated to them rises. Research in this area is, therefore, of high importance and provides valuable insight in how to prevent or counteract intrusions. Machine learning algorithms seem to be a promising answer for automated network intrusion detection, as their results often reach upwards of 99\%+ on datasets. Yet even with these results, the problem does not seem to be solved as the same models do not reach similar scores on real live network traffic. This indicates a problem with the datasets. In this work, we explored six recent network intrusion datasets and measured the quality of them from a design perspective and through a practical binary classification approach. Furthermore, we explored the need for complex classification models on these datasets, as research has shifted more towards using black-box models as opposed to white-box models. Through a literature study on the quality metrics of a dataset, we found there is a general lack of agreement amongst researchers regarding what makes a dataset good and realistic. We also found areas in which datasets are often lacking in and provide concrete advice on how to upgrade the quality of these datasets. For our practical classification approach, we built a general classification pipeline using a Random Forest. Three feature sets were tested, two of which form an ablation study to measure the effects of trigrams on the classification score. On all datasets the general classification model reaches good classification results (>90\%+) and for three datasets even reached state-of-the-art results. The ablation study yielded a positive effect of using trigrams for classification on the datasets. Our white-box approach performed on par or better than most black-box techniques. We conclude that black-box models are unnecessary in this problem context and that the techniques should shift back to white-box approaches. Next, we attempted to link the quality of dataset methodologies to the difficulty of obtaining state-of-the-art classification results. Apart from the complexity of the attack vector and benign traffic variety, we found no further properties that define this relationship. Finally, this work started with the assumption that real live network traffic is more complex and therefore more difficult to classify well on than most available datasets. Newer datasets do show improvements over older datasets and our classification results corroborated the validity of the assumption.

The field of finance is an interesting field in which much research takes place. In particular, its sub-field of modeling the dynamics of order books is an interesting field, since it translates into modeling the behaviour of traders on the market. Most of the models proposed in this field can be divided into stochastic models or machine learning models. The problem with the former ones is that they often make assumptions that do not propagate to practical applications, while the problem with the latter is that the models are often incomprehensible, even though they yield good results. This work tries to combine both of those fields to maintain the transparency of stochastic models and the accuracy of machine learning models. We target the limit order book, which is a book used in financial markets to place orders for financial securities in. It is a challenge to both generate realistic order books and predict them as well, since the model should be able to understand the underlying dynamics of order books to do this. We present a deep-learning-based model that is able to generate order books, through the use of a generative adversarial network (GAN) like architecture, and is able to predict order books using a similar architecture for the generator. Our models are able to outperform traditional generative and predictive models in both instances, while it remains transparent in its calculation: it is still able to extract useful and interpretable features from the underlying data set. These features, like the probabilities for up or down movements, or stagnations, or the rates at which the variables within the order book move, can be used by traders or algorithms to make sense out of the predictions. Traders can, for example, ask themselves whether the extracted features are representative of the analyzed asset and can, through this question, build additional confidence in the model. Another use case could be to utilize those features as inputs to other models. To our knowledge, this is the first work that proposes a combination of a stochastic and machine learning model to learn the dynamics of the order book.