@phdthesis{GhoshDastidar2025, author = {Ghosh Dastidar, Kanishka}, title = {Using Context for Credit Card Fraud Detection}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:739-opus4-15561}, school = {Universit{\"a}t Passau}, pages = {147 Seiten}, year = {2025}, abstract = {Online payment fraud is one of the biggest challenges accompanying the ubiquitous adoption of digital payment methods. The academic literature shows that providing data-driven models with additional context of a transaction results in significant improvements in fraud detection performance. However, the methods used to generate suitable context representations often rely on human expert knowledge, which is expensive and suffers from several limitations. In this thesis, we propose different methods to automate this process by learning these context representations end-to-end on the fraud detection objective. Each of these methods is evaluated on millions of real-world transactions from Worldline, our industrial partner. Central to this thesis is our proposal of the Neural Aggregate Generator (NAG), a neural network that learns context representations automatically. The architecture of the NAG is designed to resemble the structure of expert feature aggregates, while also addressing their limitations. Our evaluation of the NAG reveals that it outperforms both approaches that use expert aggregates and other end-to-end methods across several months of testing. A thorough evaluation shows that the NAG improves over other approaches on several key factors including model size and robustness to shorter sequences. We propose several extensions to the NAG with the dual motive of improved alignment with expert aggregates and improved expressiveness. Our evaluation of these extensions shows comparable performance to the NAG with ancillary benefits in terms of prospective interpretability and model size. We also introduce the novel paradigm of using \lq future' transactions as context. Our analysis of real-world data from Worldline shows that verification of transactions are often delayed by several days and that within this delay there are often several transactions booked on the card which can be used as additional context. We show that this future context improves the performance of sequence models. Moreover, we also show that a balance between past and future context yields the best results and that using future context allows the use of shorter sequences overall. Beyond context-based fraud detection, we also provide an initial proposal of generating synthetic credit card data using Generative Adversarial Networks (GANs), showing that a Wasserstein GAN can be used to generated synthetic data similar to a popular publicly available credit card fraud dataset. We also describe several possible directions for future work including the incorporation of a adapted self-attention mechanism to the NAG and the use of transformers for synthetic data generation.}, language = {en} } @phdthesis{Zwicklbauer2017, author = {Zwicklbauer, Stefan}, title = {Robust Entity Linking in Heterogeneous Domains}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:739-opus4-5047}, school = {Universit{\"a}t Passau}, pages = {iv, 191 Seiten}, year = {2017}, abstract = {Entity Linking is the task of mapping terms in arbitrary documents to entities in a knowledge base by identifying the correct semantic meaning. It is applied in the extraction of structured data in RDF (Resource Description Framework) from textual documents, but equally so in facilitating artificial intelligence applications, such as Semantic Search, Reasoning and Question and Answering. Most existing Entity Linking systems were optimized for specific domains (e.g., general domain, biomedical domain), knowledge base types (e.g., DBpedia, Wikipedia), or document structures (e.g., tables) and types (e.g., news articles, tweets). This led to very specialized systems that lack robustness and are only applicable for very specific tasks. In this regard, this work focuses on the research and development of a robust Entity Linking system in terms of domains, knowledge base types, and document structures and types. To create a robust Entity Linking system, we first analyze the following three crucial components of an Entity Linking algorithm in terms of robustness criteria: (i) the underlying knowledge base, (ii) the entity relatedness measure, and (iii) the textual context matching technique. Based on the analyzed components, our scientific contributions are three-fold. First, we show that a federated approach leveraging knowledge from various knowledge base types can significantly improve robustness in Entity Linking systems. Second, we propose a new state-of-the-art, robust entity relatedness measure for topical coherence computation based on semantic entity embeddings. Third, we present the neural-network-based approach Doc2Vec as a textual context matching technique for robust Entity Linking. Based on our previous findings and outcomes, our main contribution in this work is DoSeR (Disambiguation of Semantic Resources). DoSeR is a robust, knowledge-base-agnostic Entity Linking framework that extracts relevant entity information from multiple knowledge bases in a fully automatic way. The integrated algorithm represents a collective, graph-based approach that utilizes semantic entity and document embeddings for entity relatedness and textual context matching computation. Our evaluation shows, that DoSeR achieves state-of-the-art results over a wide range of different document structures (e.g., tables), document types (e.g., news documents) and domains (e.g., general domain, biomedical domain). In this context, DoSeR outperforms all other (publicly available) Entity Linking algorithms on most data sets.}, subject = {Linked Data}, language = {en} } @phdthesis{Dietz2011, author = {Dietz, Sebastian}, title = {Autoregressive Neural Network Processes - Univariate, Multivariate and Cointegrated Models with Application to the German Automobile Industry}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:739-opus-22524}, school = {Universit{\"a}t Passau}, year = {2011}, abstract = {Prediction of economic variables is a basic component not only for economic models, but also for many business decisions. Nevertheless it is difficult to produce accurate predictions in times of economic crises, which cause nonlinear effects in the data. In this dissertation a nonlinear model for analysis of time series with nonlinear effects is introduced. Linear autoregressive processes are extended by neural networks to overcome the problem of nonlinearity. This idea is based on the universal approximation property of single hidden layer feedforward neural networks of Hornik (1993). Univariate Autoregressive Neural Network Processes (AR-NN) as well as Vector Autoregressive Neural Network Processes (VAR-NN) and Neural Network Vector Error Correction Models (NN-VEC) are introduced. Various methods for variable selection, parameter estimation and inference are discussed. AR-NN's as well as a NN-VEC are used for prediction and analysis of the relationships between 4 variables related to the German automobile industry: The US Dollar to Euro exchange rate, the industrial output of the German automobile industry, the sales of imported cars in the USA and an index of shares of German automobile manufacturing companies. Prediction results are compared to various linear and nonlinear univariate and multivariate models.}, subject = {Nichtlineare Zeitreihenanalyse}, language = {en} }