Trace-based contextual recommendations

Nowadays, many recent applications retain traces of their usage by collecting user information. These traces help to understand user behavior and activities thus reflect the context. This paper describes how interaction traces allow the building of contextual recommendations using a Trace-Based Reasoning approach. Trace-Based Reasoning is an artificial intelligence paradigm that draws its inspiration from Case-Based Reasoning. In TBR, modeled traces act as the main knowledge container. The application considered in this paper, Wanaclip, is an interactive online video clip composition application. We added a recommendation system that guides users in both video creation and selection. The recommendation engine is fueled by interaction traces provided by previous users and is stored in TStore, a Trace Base Management System that handles the storage, processing and exploitation of traces. This approach uses similarity measures for finding and comparing episodes of traces. We validate our approach by proposing a variation to the classical accuracy definition, which we call ”acceptance rate”. Our evaluations show that this approach offers satisfactory results in term of recommendations and response time.