PTHMM: Beyond Single Specific Behavior Prediction

Existing works on user behavior analysis mainly focus on modeling a single behavior and predicting whether a user will take an action or not. However, users’ behaviors do not always happen in isolation, sometimes, different behaviors may happen simultaneously. Therefore, in this paper, we try to analyze the combination of basic behaviors, called behavioral state here, which can describes users’ complex behaviors comprehensively. We propose a model, called Personal Timed Hidden Markov Model (PTHMM), to settle the problem by considering time-interval information of users’ behaviors and users’ personalization. The experimental result on sina-weibo demonstrates the effectiveness of the model. It also shows that users’ behavioral state is affected by their historical behaviors, and the influence of historical behaviors declines with the increasing of historical time.