Research On Recommender System Based On Users' Behaviors Sequence Data

In recent years,with the rapid development of information technology,various online service platforms relying on the Internet have blossomed and flourished.At the same time,the problem of information overload has become increasingly serious.In order to reduce the time cost for users to search for interesting information and increase revenue for the platfor-m,recommender systems came into being.Ultimately,the online behaviors(such as brows,click,buy,etc.)of a user can naturally form a dynamic sequence over the timeline,thus among the recommended models,recurrent neural network(RNN)and convolutional neural network(CNN)are the two most common basic components for modeling users' behaviors sequence.However,the RNN-based recommendation methods are usually hard to preserve the long-term dependencies,and limited by the recurrent structure,these methods usually have expensive time cost for offline training and online inference.Affected by the size of the convolution kernel,the CNN-based recommendation methods tend to overemphasize the interaction between several continuous behaviors of a user,which makes it also hard for them to capture the long-term dependencies.Recently,since the attention mechanism is better at capturing the long-term dependencies and is more applicable in parallel computing,there have been some studies that applied the attention mechanism to recommender systems.However,these attention-based methods may still face some problems,such as difficulty in capturing the fine-grained preferences of a user and ignoring that a user's preferences may be multi-faceted,etc.In view of the above,this paper abandons the recurrent structure and constructs a novel neural network model,it contains two main components named item-level attention mech-anism and feature-level self-attention mechanism.Specifically,the former first picks out the behaviors related to her preferences from a user's behavior sequence,and then the latter picks out the features related to the user's preferences from these related behaviors,finally the model can capture the fine-grained preferences of the user.In addition,considering that a user's preferences include the long-term preferences and the short-term preferences,and the short-term preferences seem to be more reflective of a user's intention for the next action,thus the structure of the previous model is modified and improved.Specifically,item-level attention mechanism is adopts to capture a user's multi-faceted long-term preferences,and feature-level self-attention mechanism and vertical convolution operation are combined to capture a user's fine-grained short-term preferences.Finally,through a dimension-wise gating unit,the user's long-term preferences and short-term preferences are fused together.In this paper,the proposed model is experimentally verified on real data sets,and the results show that the proposed method can achieve a better recommendation preference.