Time-GRU Based User Behavior Sequence Modeling

As the core of personalized search and recommendation system,user behavior model-ing determines the effectiveness of the personalized systems and has always attracted the attention of researchers.After the Recurrent Neural Network(RNN)was proposed,it is increasingly popular in behavior prediction and language modeling by the reason of its ex-cellently ability in fitting sequential data.However,the traditional RNN frameworks only considers the sequential relationship of the objects and has no obvious notion of the time interval.In constrast,as for the sequence modeling problem,the behavior time interval is an important feature of the behavior relationship.Therefore,this paper improves on a widely used variant of the RNN framework,the GRU.This paper add a time-gate structure and a time-based attention mechanism,and we named it Time-GRU.These two modules are specially designed so that we can predict the users' long-term and short-term preferences to improve the recommendations and ranking results.We experiments with two well-known public datasets and an huge offline dataset on an e-commerce platform.We also conduct an AB test under the online environment.The experimental results show that Time-GRU surpasses the current state-of-the-art algorithms.