Temporal dynamic recommendation based on data imputation through association analysis

摘要

As a novel method for modelling user interest drift over time, we explore the session-based temporal dynamic recommendation, in which we impute missing rating in terms of users' association. Firstly, we mine user association groups through association analysis according to users' common preferences. Secondly, the user's consumption history is divided into sessions, and we impute vacant values based on the correlation and occupation of user association groups in each session. Thirdly, we model the user interest drift over time by latent Dirichlet allocation (LDA) in each session and predict user's current interest by an exponential decay function. Finally, we predict ratings on items for active user using neighbour-based collaborative filtering (CF). Experiments on a real dataset show that the proposed framework is more effective than previous methods on several tasks.