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.

Training machine learning (ML) models for natural language processing usually requires lots of data that is often acquired through crowdsourcing. In crowdsourcing, crowd workers annotate data samples according to one or more properties, such as the sentiment of a sentence, the violence of a video segment, the aesthetics of an image, ... To ensure quality of the annotations, several workers annotate the same sample, and their annotations are combined into one unique label using aggregation techniques such as majority voting. When the property to be annotated by the workers is subjective, the workers’ annotations for one same sample might differ, but all be valid. The way the annotations are aggregated can have an effect on the fairness of the outputs of the trained model. For example only accounting for the majority vote leads to ignoring the workers’ opinions which differ from the majority and consequently being discriminative towards certain workers. Also, ML models are not always designed to account for individual opinions, for simplicity's or performance's sake. Finally, to the best of our knowledge, no method exists to assess the fairness of a ML algorithm predicting a subjective property. In this thesis we address such limitations by seeking an answer to the following research question: how can targeted crowdsourcing be used to increase the fairness of ML algorithms trained for subjective properties' prediction? We investigate how annotation aggregation via majority voting creates a dataset bias towards the majority opinion, and how this dataset bias in combination with the current limits of ML models lead to an algorithmic bias of the ML models trained with this dataset and unfairness in the model’s outputs. We assume that an ML model able to return each annotation of each user is a fair model. We propose a new evaluation method of the ML models' fairness, and a methodology to highlight and mitigate potential unfairness based on the creation of adapted training datasets and ML models. Although our work is applicable to any kind of label aggregation for any data subject to multiple interpretations, we focus on the effects of the bias introduced by majority voting for the task of predicting sentence toxicity. Our results show that the fairness evaluation method that we create enables to identify unfair algorithms and compare algorithmic fairness, and the final fairness metric is usable in the training process of ML models. The experiments on the models point out that we can mitigate the biases resulting from majority voting and increase the fairness towards the minority opinions. This is provided that the workers’ individual information and each of their annotations are taken into account when training adapted models, rather than only relying on the aggregated annotations, and that the dataset is resampled on criteria according to the favoured aspect of fairness. We also highlight that more work needs to be done to develop crowdsourcing methods to collect high-quality annotations of subjective properties, possibly at low-cost.

In the use of Machine Learning systems, attaining the trust of those that are the end-users can often be difficult. Many of the current state-of-the-art systems operate as Black-Boxes. Errors produced by these Black-Box systems, without further explanation as to why these decisions were made, will deteriorate trust. This effect is especially strong when these erroneous decisions are generated with high confidence. This thesis presents both a data-driven as well as a human-in-the-Loop based methodology to characterize and mitigate high-confidence errors. We propose an Iterative Expert Session based methodology. By engaging domain experts through a series of interaction sessions, we aim to reduce the disconnect and knowledge gap between data scientists and domain experts, and to ultimately increase trust in the model. A practical approach was taken working in close connection with the practice of the data scientists of the ILT, helping them in improving their model and providing a direct contribution. We study the problem in the context of Road and Transportation law violations, by engaging inspectors (i.e., domain experts) in day-in-the-life and in-house interview sessions. A thorough analysis is performed of the most important features for data instances that were in error with a high degree of confidence. A method is presented that helps in characterizing these errors by predicting errors. We show that by careful removal of biased data features, proper data selection and by bridging the knowledge gap between domain experts and data scientists, we can improve the performance of the machine learning model. We show an increase of model Precision from 0.56800 with a baseline of 0.32968 to a Precision of 0.52077 with a baseline of 0.23473. Considering the baseline, this is an increase of 28.9% in Precision. We reduce biases existent in the data by reducing variables that predict on inspector practice. The magnitude of High Confidence Errors in the top 20% errors went from 0.70435 to 0.70465 showing an improvement taking into account the reduced baseline and removal of overfitted variables.

Since the introduction of the Web, online platforms have become a place to share opinions across various domains (e.g., social media platforms, discussion fora or webshops). Consequently, many researchers have seen a need to classify, summarise or categorise these large sets of unstructured user-generated content. A field related to this task is also known as opinion mining in which various applications have focused on sentiment analysis techniques to classify opinionated documents based on sentiment. More recent, researchers have focused on stance classification to classify opinionated documents based on stance in controversial debates. However, in the case of such controversial debates it would be equally interesting to know the underlying reasons behind a stance in order to truly understand a discussion. We can call these underlying reasons as perspectives. Few have focused on distilling such perspectives from text and in this research we aim to explore the use of an unsupervised model - called joint topic models - to perform the task of perspective discovery. We define perspective discovery on a controversial debate as the process of automatically finding and extracting a structured overview of perspectives from unstructured text. The aim is to quantify how well existing joint topic models can extract human understandable perspectives between and within stances for more fine-grained opinion mining on textual debates. To perform this evaluation we propose an evaluation setup with an extensive user study. This setup focuses on the topic model’s clustering ability of perspectives as well as the human understandability of the topic model’s output. Based on the results we may derive that topic models can discover some of the perspectives from text. Moreover, the results suggest that users are not influenced by their pre-existing stance when interpreting the output of topic models.

In order to combat increasing spam calls, many applications are developed and downloaded to block those calls. Some studies about performance of the applications were previously conducted, however, the actual processes the applications go through to intercept and block the spam calls are not well studied. This research presents a method to systematically explore Android APIs that are responsible for intercepting and blocking the spam calls.

The aim of this research is to provide a structured approach for dynamically analysing Android applications, focusing on applications that block or flag suspected spam caller IDs. This paper discusses ways to determine how an application stores the data regarding the phone numbers, and what APIs it utilizes. In order to develop a methodology of what to search for while performing dynamic analysis, the research in this paper focuses on analysing one such Android application, using it as a template. Additionally, it provides insights and limitations on this analysed application, and potential improvements both on the techniques and tools used.

Artificial Intelligence (AI) is increasingly helping people with all kinds of tasks, due to its promising capabilities. In some tasks, an AI system by itself will take over tasks, but in other tasks, an AI system making decisions on its own would be undesired due to ethical and legal reasons. In those cases, AI can still be of help by forming human-AI teams, in which humans get advice from the AI system helping them with making their final decisions. Human-AI teams are for instance used in the medical and legal fields. One problem arises, in which instances should one trust an AI system and in which not? Trusting the AI system when it is correct and trusting yourself when you are correct, results in a high appropriate reliance. If users appropriately rely on AI systems, it is possible to achieve complementary team performance, which is better than any single teammate. However, as known from previous literature, people struggle with assessing their performance and knowing how well they perform compared to peers. When one overestimates their performance this can be because of a dual burden, due to the lack of skill they also lack the skill to accurately estimate their performance. This phenomenon is called the Dunning-Kruger Effect (DKE). This raises questions about whether the inability to estimate their own capabilities would also reflect on their assessment of the AI system its performance. In this thesis we look at how the DKE affects (appropriate) reliance on AI systems and if so, how such effects due to the DKE can be mitigated. The effects of the DKE and possible mitigation are being tested through an empirical study (N = 249). The attempt at mitigation is done by including a tutorial intervention, which has been proved in previous research to be useful in decreasing the DKE. The tutorial intervention is aimed at revealing the weaknesses of the participant and making them aware of their miscalibrated self-estimation. Furthermore, in this thesis, the effects of revealing how the AI system makes its decisions through explainable AI (XAI) are explored. The XAI consisted of highlights from logic unit-based explanations, it should allow participants to gain more understanding of the AI advice. This thesis shows how this will affect user self-assessment and reliance on the AI system. We found that participants who overestimate themselves tend to rely less on the AI system, compared to participants that had an accurate or underestimation of their performance. After the tutorial participants have a better calibration of their self-assessment. While the designed tutorial intervention can help participants calibrate their self-assessment, it fails to promote (appropriate) reliance. Furthermore, the logic units-based explanations did not improve accurate self-assessing or increase user (appropriate) reliance on AI systems. This thesis shows the importance of considering cognitive biases when dealing with human-AI teams and invites more research on how to handle and mitigate the DKE in human-AI decision making.

Algorithmic decision-making (ADM) is becoming increasingly prevalent in society, due to the rapid technological developments in Artificial Intelligence. ADM make substantially impactful decisions about people: diagnosing whether we have a disease, what news and which ads we get to see, whether we are eligible for a job, benefits, a college or a loan, they show us personalized media and news, and steer the car that drives us home. However, ADM brings about ethical, legal and social issues by inheriting and perpetuating human biases, learning to discriminate—even learning gender or racial stereotypes, and lacking transparency and accountability. This unexpected and biased behaviour arises because these software systems are usually built without the specification of fairness requirements (i.e. what fair behaviour is expected of the system). We envision a Software Engineering for Values (SEfV) method that solves this problem. This study addresses that specification problem, aiming to help practitioners design ADM software for fairness. Using literature in social sciences—specifically organizational justice—the human value of fairness has been conceptualized in regard to ADM. This resulted in a fairness tree with four dimensions (procedural, distributive, informational and interpersonal fairness), which is further specified into 31 fairness norms. Subsequently, the fairness tree is related to current measures of fairness and techniques. Finally, we put forward the Software Engineering for Values (SEfV) framework, based on the principles of Software Engineering and Design for Values, and show how it can be applied to design ADM for fairness. Experiments were conducted where participants (N = 12) performed a design task (M = 3, 75 requirements specified) and an audit task for a hypothetical loan decision system—using a prototype of the SEfV framework. Participants found the prototype useful for both design as auditing, especially as a tool for reflecting on fairness considerations. This suggests that a high fidelity version would be useful for practitioners.

In recent years financial fraud has seen substantial growth due to the advent of electronic financial services opening many doors for fraudsters. Consequently, the industry of fraud detection has seen a significant growth in scale, but moves slowly in comparison to the ever-changing nature of fraudulent behavior. As the monetary losses associated with financial fraud continue to grow, so does the need for efficient automated decision making systems. Simple decision making rules are often still the industry standard and only show decent results in the short-term, as reverse-engineering such rules is an easy task for smart fraudsters. Supervised learning systems as automated fraud detectors have shown promising results across the field, but are plagued by challenges uniquely prevalent in the field. Disproportional class imbalance in fraudulent transactions, as well as fraudsters continually adopting new schemes make training robust and generally applicable machine learning models an arduous task. This work introduces a novel machine learning pipeline, which makes use of carefully selected synthetic samples of this minority class to augment the training dataset of the supervised model. Synthetic samples representing fraudulent transactions are filtered based on a novel technique to quantify their expected performance as an adversarial example, using both data-driven and human-expert-driven techniques. By providing the supervised model with high-quality synthetic adversarial examples, we aim to improve its generalizability to never-seen-before fraudulent behavior and, in turn, improve its robustness to the volatile nature of financial fraud. Our results show that weakly-supervised models trained on our augmented datasets are able to detect 7% more fraudulent transactions compared to a baseline model trained on the standard dataset, at the cost of a 1% increase in false positives. Our calculations further show that applying this system could lead to a decrease of 1/6 in monetary losses incurred by financial fraud.

While the performance of traditional confidence-based rejectors is heavily dependent on the calibration of the pretrained model, this study proposes the concept of feature-based rejectors and the whole pipeline where such rejector can be used in. Multiple design and development decisions along the implementation are discussed in the paper - involving humans in the loop of creating the feature set that machine should use and defining value, using interpretability methods such as saliency map to extract the features that the machine actually uses, and SceneRejector is the end product of the whole process. It is relevant to the field of computer vision, and it is applied to the task of scene classification within the scope of this paper. Its performance is evaluated against baselines with accuracy, rejection rate, and a new metric named value, which more comprehensively measures the practical value of machine learning with a reject option in different use cases. Experiments have proved the concept of a feature-based rejector works and it is able to filter out unknown unknowns, which is a challenge to the confidence-based rejectors. It also creates better value than the traditional confidence-based rejectors in some cases, especially when the machine learning model is not well calibrated. Further analysis is conducted to understand the behavior of SceneRejector as well. It is discovered that SceneRejector brings better value than confidence-based rejectors when the pretrained model is not well calibrated, and that rejection rate and accuracy of the rejector is also correlated with the value.

Machine Learning models are increasingly used to assist or replace humans in a variety of decision-making domains. However, a lot of concerns have been raised about the impact of these decisions on people’s lives. In this work we focus on two main problems. The first one is that there might be discrimination between different groups of people with respect to their protected attributes in the aforementioned Machine Learning decision making applications. The second one is that there is lack of methods that actually interpret and explain the predictions of these Machine Learning systems which are then used to help decision making. Particularly, we focus on a specific aspect of these decision making applications, namely in Machine Learning training data for the problem of classification. Most research tackles solely individual aspects like trying to detect and mitigate bias and does not pay any attention to explain their reasoning in a human interpretable way. Also, the main method that they use is to balance the distribution of the training data with respect to the protected attribute. However, as we show, this is not always the solution to the problem. On the contrary, we study concurrently three steps (detection, semantic interpretation and mitigation of bias) to overcome these shortcomings and limitations and show that there is the presence of specific visual clues that leads to that bias. Finally, we perform an extensive evaluation of our method in order to verify its efficiency and effectiveness on the particular use-case Gender Bias in Profession Prediction from Images.

Consumers share their experiences or opinion about products or brands in various channels nowadays, for example on review websites or social media. Sentiment analysis is used to predict the sentiment of text from consumers about these products or brands in order to understand the tone of customers towards these products or brands. This thesis addresses sentiment analysis in the product domain on sentence level. In this thesis three data types are used which are collected by Unilever, review data which is text that contains the opinion of a customer towards a specific product. Social data, which can be tweets, Facebook messages, Instagram messages etc. and phone data which is a summary of a phone call of a customer about a specific product. When conducting sentiment analysis one solution is to extract features from the data which can be given to a machine learning algorithm together with sentiment labels given by human annotators. The machine learning algorithm will generate a classifier which can predict a label for sentences. In sentiment analysis literature it is often not clear why certain features are chosen or for which data type certain features will work well. In this research we compare the differences when using several feature sets for the different data types. We propose three feature sets for review data and three feature sets for social data. We focus on two aspects, comparing the different feature sets and comparing the data types. In our results we do not find significant differences in performance between the feature sets. The results suggest there might be feature sets which can improve sentiment analysis specifically for the data type, but a general feature set with standard features can be comparable to that result.

In order to provide accurate statistics for industries, the classification of enterprises by economic activity is an important task for national statistical institutes. The economic activity codes in the Dutch business register are less accurate for small enterprises since small enterprises are not labelled manually. To increase the quality of the register, automatic classification of enterprises based on their websites has been tried with supervised text mining techniques. The performance of current supervised text mining techniques is limited by the available accurately labelled training data. Since inaccurate labels are available for all enterprises, the current study investigates how to leverage the noisy labelled data to improve the economic activity classification of small enterprises based on their webpage texts. The current study compares the performance of various semi-supervised methods that enlarge the training data by leveraging the abundance of noisy labelled data. The methods are compared against a supervised baseline, which uses all noisy data as is. The proposed proportional weakly self-training method queries noisy labelled instances through high probability sampling and filters mispredicted instances. Results showed that proportional weakly self-training improves upon the supervised baseline while requiring far less training instances. From qualitative analyses, we conclude that the filter of proportional weakly self-training reduces error propagation compared to classic self-training. Additional experimental results showed that large enterprises are less suitable as training data for prediction of small enterprises and that top-k performance scores improve results but are not yet sufficient for semi-automatic classification. Further examination of error detection methods is recommended to improve web-based economic activity classification.

In this thesis, we study automatically generating explanatory reports for anomalous incidents in a train control system (TCS) using Natural Language Generation (NLG). A TCS is a type of safety-critical software that allows train controllers to correctly set the tracks for a train to pass. The goal of this research is to process the majority of log files generated by such a system, detect anomalies that have occurred and represent that data in human-readable reports that explain anomalous incidents. The reports are generated by making use of NLG techniques, namely data-to-text. We design an NLG pipeline that incorporates novel graphical components. We perform all steps of the report generation process, which is the processing of data, anomaly detection and representing the data in natural language. To process the data we incorporate complex domain-specific rules which require extensive work to accommodate five separate types of event log files. Analyzing logs and detecting anomalous incidents proved complex due to the complicated structure of logs and the ill-defined relationships between different log file types. The log data explained with the NLG pipeline is more extensive than we have seen in literature, as it needs to be linked using complex domain-specific rules and includes temporal and geographical aspects of the railway system. Furthermore, the detected anomalies are very diverse and thus it is unclear what aspects should be included in a report for each anomalous case. We research whether a purely textual presentation or a combination of modalities provides higher information quality. The product of the NLG system are three presentation versions of incident reports, Text, TimeText and RailText. Each version focuses on different important aspects of the nature of the data, to explain anomalous incidents. Due to the diverse properties of the anomalous cases detected, we implement a Case-Based Reasoning (CBR) system to predict the appropriate presentation to explain each incident. We evaluate our work in two expert-based evaluation phases. The first phase evaluates the quality of the different report presentations. The second phase evaluates the feasibility of the CBR system and the quality of the reports chosen by CBR. The work in this thesis is significant as we manage to translate complex data into human-readable reports that are well received by experts and help in understanding anomalous cases. We learn that there is a difference in what presentation format provides the highest information quality of reports depending on the anomalous case being explained. Therefore we conclude that the properties of anomalous cases should be taken into account when generating anomalous reports in a TCS to ensure the highest perceived information quality of the reports. CBR is shown to perform that task well. Furthermore, we find that data familiarity of experts effects their preferences for report presentations.

Open Relation Extraction (ORE) aims to find arbitrary relation tuples between entities in unstructured texts. Even though recent research efforts yield state-of-the-art results for the ORE task by utilizing neural network based models, these works are solely focused on the English language. Methods were proposed to tackle the ORE task for multiple languages, yet these works fail to exploit relation patterns that are consistent over languages. Moreover, they require additional data to train translators, hindering efficient extension to new languages. In this work, we introduce a Language-consistent Open Relation Extraction Model (LOREM). By adding a language-consistent component to the current state-of-the-art open relation extraction model, we enable exploitation of information from multiple languages. Since we remove all dependencies on language-specific knowledge and external NLP tools such as translators, it is relatively easy to extend our model to new languages. An extensive evaluation performed on 5 languages shows that LOREM outperforms state-of-the-art monolingual and cross-lingual open relation extractors. Moreover, experiments on low- and even no-resource languages indicate that LOREM generalizes to other languages than the languages that it is trained on.