Research On Temporal Reasoning Based Hierarchical Session Perception Model For Recommendation

Session-based recommendation is defined to predict the user's next click behavior by relying only on the sequence of clicked items in the target session when the user is anonymous,which is quite common in many Internet services,such as e-commerce and short-video platform.The biggest difference between session-based recommendation and traditional recommendation is that it does not use the user's historical behavior and portrait,which is a cold-start problem,so it is difficult to model the user's interest.Existing model for session-based recommendation mostly use recurrent neural network to model the temporal relation.However,because the traditional recurrent neural network assumes that all neighbouring items in the sequence are related,when modeling the user's behavior,all items in the sequence are indistinguishable.However,the user's click pattern may be variable and unstable.Therefore,the traditional recurrent neural network may learn the wrong dependencies between the items in the session.In addition,the existing session-based recommendation model ignores the rich and diversified features of the items in the session when modeling the user's interest,which is not conducive to the modeling of user interests.In order to solve these two problems,we propose a temporal reasoning based hierarchical session perception recommendation model(TRHSR).The main contributions of TRHSR are:1)We propose a method to model user's behavior.By introducing relational memory core(RMC),we allow each time step in the session to interact with each other,thereby learning complex dependencies in the session,effectively avoid learning wrong dependencies.2)We propose a method to model user's interest.On the basis of the temporal reasoning mechanism,a hierarchical attention mechanism is proposed to model user's interest at two levels,namely feature-level and instance-level,and amplify the important features to the downstream,for modeling user's interest in a fine-grained way.We conduct a series of experiments on two public datasets,comparing our model with baseline model and advanced model.The results show that TRHSR is superior to baseline model and advanced model on both datasets,which proves the effectiveness of the proposed TRHSR model.