Ti-SSE: a time interval based sequential recommendation model via stochastic shared embeddings

Abstract

Sequential recommendation systems attempt to use the sequence of user interactions to predict users' next behavior according to their recent based on their most recent actions. Markov chains, recurrent neural networks and self-attention have been widely developed because of their traditional ability to capture the dynamics of sequential patterns. However, most of these models make a simplified assumption that they treat the interaction history as an ordered sequence and ignore the time interval between each interaction. In other words, their model is a time series without considering the actual time interval. In this paper, we try to explicitly model the interactive timestamp in a sequence model framework and explore the impact of different time intervals on the prediction of the next project. We propose the model Ti-SSE which combines TiSASRec and SSE-PT model. Thus, we avoid the shortcomings of the traditional sequential recommendation models that ignore the absolute time intervals while the Stochastic Shared Embeddings (SSE) regularization makes our model far from overfitting due to excessive use of self-attention blocks. The results on several datasets show that our model Ti-SSE has advantages over the previous models.