Contributions published at Dynamic and Distributed Information Systems (Abraham Bernstein)

Knowledge graph embedding aims to learn distributed representations for entities and relations, and are proven to be effective in many applications. Crossover interactions --- bi-directional effects between entities and relations --- help select related information when predicting a new triple, but hasn't been formally discussed before. In this paper, we propose CrossE, a novel knowledge graph embedding which explicitly simulates crossover interactions. It not only learns one general embedding for each entity and relation as in most previous methods, but also generates multiple triple specific embeddings for both of them, named interaction embeddings. We evaluate the embeddings on typical link prediction task and find that CrossE achieves state-of-the-art results on complex and more challenging datasets. Furthermore, we evaluate the embeddings from a new perspective --- giving explanations for predicted triples, which is important for real applications. In this work, explanations for a triple are regarded as reliable closed-paths between head and tail entity. Compared to other baselines, we show experimentally that CrossE is more capable of generating reliable explanations to support its predictions, benefiting from interaction embeddings.

In major political events and discussions, recommender algorithms play a large role in shaping opinions and setting the agenda for more traditional news media. These algorithms are used pervasively in social networks, media platforms and search engines. They determine what information is shown to the user and in which order. Existing recommender systems focus on improving accuracy based on historic interaction-data, which has received criticism for being detrimental to the goals of improving user experience and information diversity. Their accuracy is measured in terms of predicting future user behavior based on past observation. The most popular algorithms suggest items to users based on the choices of similar users. However, these systems are observed to promote a narrow set of already-popular items that are similar to past choices of the user, or are liked by many users. This limits users' exposure to diverse viewpoints and potentially increases polarization. One reason for the lack of diversity in existing approaches is that they optimize for average accuracy and click rates. There are other aspects of user-experience like new information and diering viewpoints, that are not taken into account. Similarly, a high average accuracy does not necessarily mean that the algorithm is similarly accurate for dierent groups of users and niche items. Besides this, training data may be biased because of previous recommendations, or the data collection process. Although information diversity is important, its denition depends heavily on the specic application and it is hard to nd a general denition in the context of recommender systems. Nevertheless, there are denitions that pertain to some specic aspects of diversity, such as entropy, coverage, personalization, and surprisal. There is a need to design recommender algorithms that balance accuracy and diversity in order to deal with the aforementioned problems. To approach this problem, we rst investigated state-of-the-art algorithms and found that they promote already popular products. We also found that recommender diversity depends on the optimization parameters of the algorithms, a factor that had not received enough attention before. In this context, we developed a graph- based method that promotes niche items without sacricing accuracy. We also developed a probabilistic latent-factor model that signicantly improves the coverage and long-tail diversity of recommendations. Currently, we are using these insights to develop new machine learning algorithms for diverse news recommendation. We plan to and deploy and test them in a large-scale experiment involving multiple news producers and consumers.

Analyzing data streams is a vital task in data science. Often, data comes in different shapes such as triples, tuples, relations, or matrices. Traditional stream processing systems, however, only process data in one of these formats. To enable the processing of streams combining different shapes of data, we developed a system that parses SPARQL queries using the Apache Jena parser and transforms them to Apache Flink topologies. With a custom data type and tailored functions, we enabled the integration of matrices in Jena and therefore allowed to mix graphs, relational, and linear algebra in an RDF graph. This provided a proof of concept that queries may be written for static data and – with the usage of the streaming engine Flink – can easily be run on data streams, even if they contain multiple of the aforementioned types.

Knowledge Bases (KBs) and textual documents contain rich and complementary information about real-world objects, as well as relations among them. While text documents describe entities in freeform, KBs organizes such information in a structured way. This makes these two information representation forms hard to compare and integrate, limiting the possibility to use them jointly to improve predictive and analytical tasks. In this article, we study this problem, and we propose KADE, a solution based on a regularized multi-task learning of KB and document embeddings. KADE can potentially incorporate any KB and document embedding learning method. Our experiments on multiple datasets and methods show that KADE effectively aligns document and entities embeddings, while maintaining the characteristics of the embedding models.

Fitness trackers and apps provide a new way to monitor the behaviour and health of a person. These technologies support psychiatrists in the treatment of patients. This thesis presents a prototype that aims to identify stressful situations with a fitness tracker and an app. The prototype will help medical professionals to evaluate whether a mobile healthcare system is accepted by patients. In addition to collecting physiological data, subjective perceptions of stress and well-being can be recorded via the app. Authorized specialists can view the gathered data in a web application.

A common issue in recommender systems is the lack of explicit feedback. In this thesis we build a news recommendation framework to collect explicit and implicit feedback from users reading news articles. Our goal is to find a generic way how to interpret implicit feedback to approximate a users interest in an article. After conducting a field study we train machine learning models on the collected implicit feedback to predict the explicit feedback the user has assigned to a given article. Our results reveal that the time spent reading and reading progress are the two most influential features. These results have limitations due to UI issues with the way we collected explicit feedback in our iOS client. We realized that our design restricts the number of negative explicit feedback which lead to a severely imbalanced data set. We conclude with suggestions to improve the UI design and recommend a more comprehensive experiment to collect an overall bigger data set.

null

Word embeddings changed the possibilities in the field of Natural Language Processing and Machine Learning completely, opening new doors for many applications. One is the creation of document embeddings with the Doc2Vec algorithm based on Word2Vec. These dense distributed latent vectors allow to work with text in a better, more meaningful way compared to older text vectorization processes such as Bag-of-Words (BOW). In this thesis, a variety of baseline methods are compared in different categories against Doc2Vec. To finally asses the usefulness of these older approaches after the recent upshake in the Natural Language Processing field. Empirical results show that BOW used with a strong classifier is especially in smaller datasets better than Doc2Vec. Additionally, an approach to reduce the dimensionality of a BOW is presented.

In machine learning, embeddings are used to encode information in a vector space. Word2Vec is a popular method for creating word embeddings, vector representations of words, and can be used for semantic tasks such as finding the similarity between words. Similarly, image embeddings are vector representations of images. The concatenation of word and image vectors is one possible multi-modal model and has shown to outperform the individual models. This work examines if and when a concatenation is beneficial and proposes an alternative model without using vector concatenation.

The last years have witnessed an increasing shortage of data experts capable of analyzing the omnipresent data and producing meaningful insights. Furthermore, some data scientists mention data preprocessing to take up to 80% of the whole project time. This paper proposes a method for collaborative data analysis that involves a crowd without data analysis expertise. Orchestrated by an expert, the team of novices conducts data analysis through iterative refinement of results up to its successful completion. To evaluate the proposed method, we implemented a tool that supports collaborative data analysis for teams with mixed level of expertise. Our evaluation demonstrates that with proper guidance data analysis tasks, especially preprocessing, can be distributed and successfully accomplished by non-experts. Using the design science approach, iterative development also revealed some important features for the collaboration tool, such as support for dynamic development, code deliberation, and project journal. As such we pave the way for building tools that can leverage the crowd to address the shortage of data analysts.

null

Die föderale Schweiz eignet sich hervorragend, um experimentelle Pionierarbeit bei der Digitalisierung der Demokratie zu leisten. Warum sich das lohnt.

Inspired by the World Wide Web, the Web of Data is a network of interlinked data fragments. One of the main advantages of the Web of Data is that all of its content is processable by machines. However, this also has its drawbacks when it comes to monetization of the content: advertisements and donations—two important financial motors in the World Wide Web—do not translate into the Web of Data as they rely on exposing the user to advertisement/call for donations. The remedy this situation, we propose two different monetization strategies for the Web of Data. The first strategy involves a marketplace where users can buy data in an integrated way. The second strategy allows third parties to promote certain data. In return, the sponsors pay money whenever a user follows a link contained in the sponsored data. We identified two different kind of data—commercial and sponsored data—which can benefit from the two respective monetization strategies. With our work, we propose solutions to the problem of financing the creation and maintenance of content in the Web of Data.

The World Wide Web is a massive network of interlinked documents. One of the reasons the World Wide Web is so successful is the fact that most content is available free of any charge. Inspired by the success of the World Wide Web, the Web of Data applies the same strategy of interlinking to data. To this point, most of data in the Web of Data is also free of charge. The fact that the data is freely available raises the question of financing these services, however. As we will discuss in this paper, advertisement and donations cannot easily be applied to this new setting. To create incentives to subsidize data providers, we propose that sponsors should pay the providers to promote sponsored data. In return, sponsored data will be privileged over non-sponsored data. Since it is not possible to enforce a certain ordering on the data the user will receive, we propose to split up the data into different batches and deliver these batches with different delays. In this way, we can privilege sponsored data without withholding any non-sponsored data from the user. In this paper, we introduce a new concept of a delayed-answer auction, where sponsors can pay to prioritize their data. We introduce a new model which captures the particular situation when a user access data in the Web of Data. We show how the weighted Vickrey-Clarke-Groves auction mechanism can be applied to our scenario and we discuss how certain parameters can influence the nature of our auction. With our new concept, we build a first step to a free yet financial sustainable Web of Data.

The linking of correspondent entities between multiple knowledge graphs (KGs) is known as entity alignment. This thesis introduces the embedding-based method Dependent Learning of Entity Vectors (DELV) for entity alignment. In an iterative fashion, the method learns a low-dimensional vector representation for the entities in a satellite model in dependence of a pretrained central model. Word2vec and rdf2vec constitute the basis for the embedding learning process. DELV is evaluated on real-world datasets, originating from the three knowledge graphs DBpedia, Wikidata and Freebase. DELV outperforms most of its baselines in terms of the mean rank, the hits@1 and hits@10. While entity alignment is normally performed on two KGs, this thesis also demonstrates how DELV can be efficiently used for alignment of unlimited KGs.