Step-Enhancement of Memory Retention for User Interest Prediction

User behavior modeling and interest prediction are always the key elements in preference analysis, product recommendation and personalized service. The psychological memory theory has been proved capable of reflecting changes in user interest. However, merely focusing on the memory forgetting mechanism (the retention-only model) or considering the superimposition of interest retention in short term (the gradual-retention model), existing methods have a poor prediction capability because of ignoring the long-term impact of repeated behaviors. In this paper, we propose a step-enhancement of memory retention (SEMR) model which integrates the cross-enhancement-effects of multiple historical behaviors under different time windows to characterize user interest. In addition, we use some extended correction methods to eliminate the effect of discontinuous records. Numerical experiments using real TV viewing data validate the efficiency of our proposed model and methods, which reduce the average prediction error to 0.3, outperforming the traditional models by around 50%.