Year: 2020
Type: dissertation
Abstract: Relation prediction seeks to predict unknown but potentially true relations by revealing missing relations in available data, by predicting future events based on historical data, and by making predicted relations retrievable by query. The approach developed in this thesis can be used for a wide variety of purposes, including to predict likely new friends on social networks, attractive points of interest for an individual visiting an unfamiliar city, and associations between genes and particular diseases. In recent years, relation prediction has attracted significant interest in both research and application domains, partially due to the increasing volume of published structured data and background knowledge. In the Linked Open Data initiative of the Semantic Web, for instance, entities are uniquely identified such that the published information can be integrated into applications and services, and the rapid increase in the availability of such structured data creates excellent opportunities as well as challenges for relation prediction. This thesis focuses on the prediction of potential relations by exploiting regularities in data using statistical relational learning algorithms and applying these methods to relational knowledge bases, in particular in Linked Open Data in particular. We review representative statistical relational learning approaches, e.g., Inductive Logic Programming and Probabilistic Relational Models. While logic-based reasoning can infer and include new relations via deduction by using ontologies, machine learning can be exploited to predict new relations (with some degree of certainty) via induction, purely based on the data. Because the application of machine learning approaches to relation prediction usually requires handling large datasets, we also discuss the scalability of machine learning as a solution to relation prediction, as well as the significant challenge posed by incomplete relational data (such as social network data, which is often much more extensive for some users than others). The main contribution of this thesis is to develop a learning framework called the Statistical Unit Node Set (SUNS) and to propose a multivariate prediction approach used in the framework. We argue that multivariate prediction approaches are most suitable for dealing with large, sparse data matrices. According to the characteristics and intended application of the data, the approach can be extended in different ways. We discuss and test two extensions of the approach--kernelization and a probabilistic method of handling complex n-ary relationships--in empirical studies based on real-world data sets. Additionally, this thesis contributes to the field of relation prediction by applying the SUNS framework to various domains. We focus on three applications: 1. In social network analysis, we present a combined approach of inductive and deductive reasoning for recommending movies to users. 2. In the life sciences, we address the disease gene prioritization problem. 3. In the recommendation system, we describe and investigate the back-end of a mobile app called BOTTARI, which provides personalized location-based recommendations of restaurants.
DOI: 10.5282/edoc.26228
Author: Huang, Yi
Institution: Ludwig-Maximilians-Universität München
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3043810141