Sequential Recommendations Based On Modeling And Predicting User Preferences With Multiple Item Attributes

With the massive growth of data on the Internet,information overload has seriously hindered the development of human society.Recommendation system should be applied to solve the problem of information overload by means of filtering,screening and matching.The core problem of traditional recommendation system is to model the interaction between users and projects according to the historical feedback of users,but this modeling method is static and can only capture the common preferences of users.In real life,users tend to be serialized behaviors rather than independent interactions.In addition,users’ preferences and the popularity of items also change dynamically.Different contexts often lead to different user project interactions.Because of its suitability for practical application scenarios,sequential recommendation has become the focus of the entire deep learning community.Although there are more and more researches on sequential recommendation system in recent years,there are still some shortcomings in exploring and using item attribute relationship to improve prediction accuracy.In this study,a new technical framework,MIA-SR,is proposed to implement sequence recommendation(SR)by modeling and predicting user preference multi-item attributes(MIA).(1)Inspired by the machine translation task,the user-project interaction is modeled in the form of sequence when modeling the dynamic behavior of users.Specifically based on Transformer architecture,feature extraction and encoding of user sequence are realized through multi-head attention mechanism,feedforward network,residual network,layer normalization,etc.(2)In order to better depict users’ various interests,in addition to using item sequence,attribute sequence is also used to generate multiple vector representations of users’ interests,and a self-attention module for interest aggregation is established to realize the fusion of user interest representations.(3)Considering that items and attributes are highly correlated and promoted semantically,graph convolutional network is further proposed to enhance the representation of items and attributes on item-attribute bipartite graphs.(4)Finally,multiple facets of user interest prediction can be regarded as multiple tasks,and independent representation and prediction results can be generated for each task.Combined with automatic learning weight loss function,the model is optimized.A large number of experiments have been conducted on common reference datasets to verify the superiority of mi A-SR over existing methods.Compared with SASRec,the MIA-SR model improves at least 2.0% HR@5-10 and 3.1% NDCG@5-10 on JD and Tao Bao datasets.The prediction accuracy was also improved in the ML-1M dataset with only two tasks.Finally,the validity of each component in MIA-SR and the expansibility of the model were verified by ablation experiment and parameter analysis.