Preview

PDF, English - main document Download (2MB) | Terms of use

Abstract

Events are ubiquitous in real-life. With the rapid rise of the popularity of social media channels, massive amounts of event data, such as information about festivals, concerts, or meetings, are increasingly created and shared by users on the Internet. Deriving insights or knowledge from such social media data provides a semantically rich basis for many applications, for instance, social media marketing, service recommendation, sales promotion, or enrichment of existing data sources. In spite of substantial research on discovering valuable knowledge from various types of social media data such as microblog data, check-in data, or GPS trajectories, interestingly there has been only little work on mining event data for useful patterns.

In this thesis, we focus on the discovery of interesting, useful patterns from datasets of events, where information about these events is shared by and spread across social media platforms. To deal with the existence of heterogeneous event data sources, we propose a comprehensive framework to model events for pattern mining purposes, where each event is described by three components: context, time, and location. This framework allows one to easily define how events are related in terms of conceptual, temporal, and spatial (geographic) relationships. Moreover, we also take into account hierarchies for contexts, time, and locations of events, which naturally exist as useful background knowledge to derive patterns at different levels of abstraction and granularity. Based on this framework, we focus on the following problems: (i) mining interval-based event sequence patterns, (ii) mining periodic event patterns, and (iii) extracting semantic annotations for locations of events. Generally, the first two problems consider correlations of events whereas the last one takes correlations of event components into account.

In particular, the first problem is a generalization of mining sequential patterns from traditional data, where patterns representing complex temporal relationships among events can be discovered at different levels of abstraction and granularity. The second problem is to find periodic event patterns, where a notion of relaxed periodicity is formulated for events as well as for groups of events that co-occur. The third~problem is to extract semantic annotations for locations on the basis of exploiting correlations of contexts, time, and locations of events. For the three problems above, we respectively propose novel and efficient approaches. Our experiments clearly indicate that extracted patterns and knowledge can be well utilized in various useful tasks, such as event prediction, semantic search for locations, or topic-based clustering of locations.