Towards semantic association rules mining from ontology-based semantic trajectories

Abstract

Different technologies and social-cultural aspects of our lives have allowed the acquisition of people's mobility data. The same applies to other moving objects, such as birds with GPS trackers and hurricanes with real-time satellite data. Although these raw positioning and timings are useful in many applications, it has been long recognized by the Trajectory Data community that semantics are required to capture the complexity of humans' and other objects' behaviors. Semantic Trajectories were proposed in this context as raw trajectories enriched with semantic annotations and possibly interlinked with external data. Based on these requirements, many works incorporate concepts and technologies from the Semantic Web to deal with the complexity of merging, representing, and querying heterogeneous data. They usually use ontologies to represent and manipulate concepts such as Moving Objects, Trajectories, Stops and Moves, and semantic aspects related to each of them. Nonetheless, we find that no previous work has explored mining patterns from these ontology-based representations. On the contrary, current efforts use standard association rule mining algorithms, such as Apriori, which require propositional data represented as Boolean feature vectors. To mine patterns aware of the semantic relations in a Semantic Trajectory ontology, we explore algorithms from the Knowledge Base Refinement field. These methods were proposed to use real-world facts represented in Knowledge Bases such as YAGO and DBPedia to infer new entities and relationships. We build on previous works describing ontology-based trajectory representations and tackle the knowledge discovery task using AMIE, a well-known state-of-the-art KB rule mining algorithm. This approach mines patterns in the form of Horn rules, which allows us to investigate associations between time, spatial, and semantic relations interlinking trajectory events. We show that representations previously proposed in the Semantic Trajectory community are not suitable to be directly mined by this approach. However, they can be easily extended to power the AMIE algorithm. We also describe and address different issues that arise when using a domain-agnostic mining algorithm. The proposed data pipeline mines interesting patterns in experiments using Foursquare datasets. Nonetheless, there is a large number of rules which state facts that are too general. We build on these issues and argue in favor of the design of a domain-specific mining algorithm. We discuss future opportunities based on the acquired experience and experiments. Our approach shows how the Semantic Trajectory and Knowledge Base Refinement communities have built in recent years a large number of representations and mining approaches that could be put together to mine rules with rich semantic expressiveness from semantic data.