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.

With advancements in Internet and technology, it has become increasingly easy for people to enjoy music. Users are able to access millions of songs through music streaming services like Spotify, Pandora, and Deezer. Access to such large catalogs created a need for relevant song recommendations. Music recommender systems assist users in finding the most relevant songs by consistently matching them with the user’s preference. Accurately representing these preferences is essential to creating accurate and effective song recommendations. User preferences are highly subjective in nature and change according to context (e.g., music that is suitable for running is not suitable for relaxing). Preferences for songs can be based on characteristics of high level audio features, such as tempo and valence.
This thesis proposes a new contextual re-ranking algorithm, which belongs to the group of contextual post-filtering techniques, to leverage users’ contextual information. The algorithm uses two models, a global and personalized model, to model user preferences. These models use audio features to represent user preference in specific contextual conditions. The algorithm is able to re-rank any given music recommendation list. First, we analyze the correlation between audio features and contextual conditions. This analysis shows that the correlations are significant, thus audio features are suitable for representing user preference in contextual conditions. Thereafter, we implement and evaluate the re-ranking algorithm using accuracy metrics on the #NowPlaying-RS and InCarMusic datasets, using various initial recommender algorithms. Results show there
is merit in applying such a re-ranking algorithm to increase recommendation accuracy. The personalized model, given enough historical data, consistently outperforms the global model.

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.

While voice assistants have exploded in popularityover the last decade, they still have many issues.Among these is the issue of result presentation:how do you speak results to the user? Priorresearch has investigated how cognitive loadrelates to result presentation and other methodsfor result presentation, but highlighted a lackof research about result presentation over speechand called for further investigation. [12] Thisresearch answered that call, and investigated the research question How can web search results bepresented over a speech-only channel such thatthe users’ cognitive load is minimized?Wetested three methods: 1) one-source, one-shotresponses, 2) question-and-answer responses, and3) multi-source, one-shot responses. In all cases,single-source responses were preferred to multi-source responses. Of the two types of single-source responses tested (question-and-answer styleresponses and one-shot responses), both types hada similar average score. Our recommendationto developers of conversational search technologywould be to allow user choice between these tworesponse types. We would only recommend multi-source responses as a fall-back option in the caseof a low estimated probability of result relevance.

For real-world problems even the most complex machine learning models can only achieve a certain accuracy. This makes it important to understand why a specific prediction is made. Explanations can provide human decision support by allowing human experts to assess the reasoning of the model as well as the correctness. Specifically, in this thesis, we consider the problem of predicting violations on inland ships in the Netherlands to help inspectors in the Dutch government
deciding which ship to inspect. The main contribution is determining confidence in a prediction separately from probability and using this confidence estimation for deciding which prediction of violation to select as well as to explain.
With the limited number of inspectors and a large number of inland ships in the Netherlands, the global performance on all ships is less relevant. Instead, deciding the most qualitative predictions is more useful. Therefore, a measure of model confidence is determined to improve upon the traditional ranking based on probability. In the evaluation of this approach, no significant difference is found between the ranking based on probability for complex ensemble models. However,
for simpler, more interpretable models, there is a significant improvement in using model confidence to re-rank. The determination of confidence is further used to create explanations from the context of confidence. The goal of these explanations is to help an inspector in deciding whether to inspect an inland ship. This novel explanation approach justifies the confidence in a prediction by expressing features contributing towards the confidence. We perform a human-grounded user study evaluation to identify the task effectiveness, perceived usefulness and user trust compared to the explanations from the traditional context of probability. The results of the user study suggest the explanations of the confidence to be particularly useful for problems with a lower accuracy.

People like to travel in groups to visit places. Group recommendation systems can be used to recommend an itinerary of "places of interests" (POIs) in an ordered sequence. The order of POIs in the sequence can be explained to group members to increase acceptance of the recommended items. There is a possibility that explanations which reveal names and rating preferences could create a threat to privacy. The main study in this work uses two group types - a primary group consisting of closely-related members, and a secondary group consisting of loosely-related members. Explanations with either complete information or privacy-preserving information are offered alternatively to these groups. The purpose of this study is to evaluate whether different group types need different types of explanations to improve their satisfaction. These explanations explain the entire recommended sequence of POIs with regard to possible conflicting situations that could occur due to disagreement with the order of the sequence. A total of 25 participants took part in the evaluation. There was no significant difference identified between the explanation types preferred by each group type. To understand the underlying reason for this result, a post-hoc analysis was done. We identified a participant's most frequently used conflict-handling modes using the Thomas-Kilmann personality assessment test. We then analyzed the user comments provided during the questionnaire. The analysis potentially suggests that different conflict-handling modes could be a factor affecting which explanation type was preferred by a person when they are in a particular group (e.g. primary vs secondary).

With the advent of Internet and resulting data boom, Recommender Systems have come to rescue by filtering the information available on the internet by providing us with relevant information. These systems come handy when one wants to listen to songs, watch movies or even buy products on the Internet. Primarily, these recommender systems used content based or collaborative filtering techniques to recommend items. More recent research has studied the importance of contextual features in recommender systems. Music preference has always been associated with the contextual feature emotion. However, few studies study the mood congruence effect in the domain of music recommender systems. The field of music emotion recognition also remains unexplored with recommendations being made with limited features. \\

This master thesis analyses the relationship between few latent musical features and user emotion through our interface MooDify. It is a music recommendation system that incorporates emotion in a user using emotion induction techniques and investigate the effect of their emotional state on satisfaction and unexpectedness when presented with songs curated to specific musical features. To achieve this, we analysed the enjoyment and unexpectedness ratings for recommendations specific to latent musical features for a given emotional state. We have been able to achieve some interesting results through this study which has been discussed later in this work.

News is the main source of information about events in our neighborhood and around the globe. In an era of digital news, where sources vary in quality and news spreads fast, it is critical to understand what is being consumed by the public. Partisan news is news that favors certain political parties or ideologies. It is undesirable because news organization should aim for an objective and balanced reporting. An automatic system that can classify news to be partisan or non-partisan is thus desired. Such a system (partisan news detector) requires a sufficient amount of data to learn the pattern of partisanship in news articles. However, these labels are expensive to collect since manual inspection of each article is required.

Inferring partisanship labels from the partisanship of publishers is another approach that has been used in previous research. By treating all articles by partisan publishers to be partisan news and those by non-partisan publishers to be non-partisan, it is easy to collect a large number of labeled articles. This way of deciding labels is noisy, making it more difficult for a detector to learn.

This thesis compared the performance of using publisher-level labels and article-level labels for partisan news detection. The detector was designed as a binary classifier. We compared the difference in performance across several feature sets to ensure the observed difference was due to the annotation level, not the choice of specific classifiers. The experiments were performed on two datasets of different properties to ensure the generalizability of the results. We found that classifiers trained with publisher-level labels have higher recall but lower F1-score compared to classifiers trained with article-level labels. We also observed that the classifiers overfit on publishers but reducing the overfitting with feature selection did not improve the performance. Comparing the performance difference between the two datasets, we concluded that an important factor that determines the performance achievable by the publisher-level labels is the quality of publishers that are included in the dataset. This is valuable for future dataset collection.

Our work provides a benchmark performance of publisher-level labels, which can be used as baselines for future research that investigate other methodologies to utilize the noisy labels. We also compared the performance between the two levels of annotation and concluded that partisan news detectors trained with article-level labels are more practical to be used in a fully-automated system since they have on average 10\% higher F1-scores than those trained with publisher-level labels. However, the high recall of the latter makes them applicable in use cases where high recall is desired.

The Natural Language Generation field has advanced in generating human readable reports for domain experts in various fields. Nevertheless, Natural Language Generation and anomaly detection techniques have not been used in the rail domain yet. Currently, data analysis and incident reporting for log files from the train control system are performed manually which is very time consuming task that is prone to missing crucial information. The rail domain is safety critical domain where detailed analysis of the train control system may prevent incidents from happening as well as help improve the performance of the train control system. This research designs, implements and evaluates a Natural Language Generation model that successfully translates anomalies detected in log files into human readable reports.
This thesis presents the steps taken for developing a Natural Language Generation system in the rail domain. Additionally, we examine two representations of the train control system used for the Content Determination task of the Natural Language Generation system. Through a case study with domain experts, we evaluate the performance and preference between the reports generated based on the two representations of the train control system and the data retrieved from the log files. The goal is to find a representation that presents the used with a full/solid understanding of the anomalies detected in the log files.
Based on the case study performed to evaluate the system, we present the finding that when developing a Natural Language Generation system for the rail domain, reports generated using a more detailed representation of the train control system (more precisely, using both state names and state attributes that specify the step by step process of setting a route for a train) were preferred over the reports generated using a less detailed representation (only state names). The preference was based on readability, accuracy and understandability measures of the reports presented during the case study.

In today’s digital world, users are often confronted with an abundance of information. Whether the user is looking to compare online prices for products, searching for new movies to watch or music to listen, the available information at hand exceeds the amount of information which the user wants to consider before making a choice. For this, various recommendation systems have been developed. Primarily, the recommendations made by these recommendations systems have been evaluated based on their accuracy. More recent research has begun evaluating the subjective perception of the recommendations and has developed additional attributes such as diversity and novelty to evaluate recommendation systems. However, the impact of such attributes on actual user satisfaction has been explored less. Recent research has seen an increase in evaluating other aspects of recommendation systems such as recommendation interfaces. Nonetheless, very limited work has been done in terms of recommendation system interfaces to improve the perceived quality of the recommendations and overall user satisfaction.

This master thesis introduces and evaluates a novel interface, MovieTweeters. It is a movie recommendation system which incorporates social information with a traditional recommendation algorithm to generate recommendations for users. Few previous studies have investigated the influence of using social information in interactive interfaces to improve the novelty of recommendations. To address this gap, we investigate whether social information can be incorporated effectively into an interactive interface to improve recommendation novelty and user satisfaction. Our initial results suggest that such an interactive interface does indeed helps users discover more novel items. Also, we observed users who perceived that they discovered more novel and diverse items reported increased levels of user satisfaction. Surprisingly, we observed that even though we successfully were able to increase the system diversity of the recommendations, it had a negative correlation with users perception of novelty and diversity of the items highlighting the importance of improved user-centered approaches.

Current research on personality and diversity based Recommender Systems (RecSys) are mostly separated. In most diversity-based Recommender Systems, researchers usually endeavored to achieve an optimal balance between accuracy and diversity while they commonly set a same diversity level for all users. Different diversity needs for users with different personalities are rarely studied. Another branch of research on personality-based Recommender Systems mostly emphasize utilizing personality information to enhancing the rating prediction accuracy so as to solve the ’Cold-Start Problem’. While few of them have in depth investigated whether and how it influences users’ other preference needs (such as diversity needs).

This thesis presents the work how we combine these two branches of research together. Anchored in the music domain, we investigate how personality information can be incorporated into the Music Recommender Systems to help adjust the diversity degrees for people with different personalities. We first conducted a pilot study to investigate the correlation between users’ personality factors and their diversity needs on the music recommendations. Results showed that there exits significant correlations between them, especially when we consider the personality factor ’Emotional Stability’. Based on such findings, we then proposed a personality-based diversification algorithm to help enhance the diversity adjusting strategy according to people’s personality information in music recommendations. Our offline and online evaluation results demonstrated that our proposed method is an effective solution to generate personalized recommendation lists with relatively higher diversity.

In recent years, personalized recommender systems have been facing criticism in research due to their ability to trap users in their circle of choices, called "filter-bubble", thereby limiting their exposure to novel content. In solving the issue of filter-bubble, past research has focused on providing explanations to users about how a recommender system recommends a specific item. This thesis addresses the issue of filter bubbles by helping users understand not just why a recommendation was made, but to also convey something about the limits of this recommendation.

In this thesis, we help users to better understand their consumption profiles by exposing them to their unexplored regions, thereby indirectly nudging them to diverse exploration. We refer to these unexplored regions as the user's blind-spots, and we visualize these by enabling comparisons between a user's consumption pattern with that of other users of the system. We compare the effectiveness of two visualizations -- a bar-line chart and a scatterplot --- in representing this consumption information and in increasing a user's intention to explore new content.

We performed a live user study to test our system (n=23). The results suggest that users are able to better understand their profile with both the visualizations.
Furthermore, our results confirmed that users with a higher understanding of their profile tend to explore their blind-spot categories more. From our experiment, we provided a first step towards increasing user's awareness of their choices as well as providing the kind of user control that encourages users to explore new types of items.

Memes are theorized to be the building blocks of culture. Due to a lack of empirical validation, however, the theory of memes — memetics — remains in its infancy. We argue that one of the missing components for such empirical validation is a method for the large-scale identification of memes.

In this thesis, we develop a method for the identification of scientific memes — ngrams of length 1 through 4, denoting scientific concepts — propagating within online communities. With data extracted from science-oriented correspondence extracted from five communities on the online discussion platform Reddit, and five communities on the online question and answer platform StackExchange, we perform a large-scale automated evaluation in which we find that memes identified in these communities correspond to the titles of Wikipedia articles; and a small-scale human evaluation in which we find that the identified memes represent relevant concepts to the community’s scientific field.

Furthermore, we introduce a slight adaptation of this method to elucidate one of memetics’ predictions: the occurrence of interactions between memes, where the occurrence of one meme has a positive or negative influence on the propagation of another meme. To evaluate this method for the identification of meme interactions, we construct meme interaction networks, in which we find that the most central memes correspond to the most relevant scientific concepts.

We find that our methods are able to extract key concepts within online communities, identifying thousands of relevant concepts from millions of candidate ngrams. Thus, our method may contribute to contemporary text mining research, and could be used in place of, or in conjunction with current approaches, such as TF-IDF or LDA.

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 an increasing urbanizing world the need for efficient transportation methods is growing. Ridesharing, sharing a taxi trip with multiple passengers, has been proposed as an effective way to contribute towards solving the traffic problems that arise in city centers. Current solutions to the problem are often more focused on optimality, and a fast algorithm is required to handle large quantities of passengers in real-time applications. We propose an algorithm that combines fast and efficient heuristics with clustering and parallelization techniques that is able to solve large problem instances within several seconds. The small response time of the algorithm is able to provide feedback to the customer rapidly, which makes it very suitable for ridesharing in real-time.
A segmentation of the customers is implemented to speed up the search. This method divides the customers into subareas based on pickup location such that each subproblem contains a similar number of customers. This division also bounds the size of each subproblem to ensure a solution can be found in real-time.
The primary search strategy uses a combination of effective local search heuristics to improve solutions for these subproblems rapidly. A technique reminiscent of variable neighborhood search is also implemented for diversification that can solve smaller instances to a greater degree. The full approach can handle problems containing over 5,000 requests and can effectively manage on-demand ridesharing in Manhattan where the current demand for yellow taxis has tripled.