Session Sequence Recommendation Based On Deep Learning

In today’s big data era,the problem of information overload is becoming more and more serious.Session sequence recommendation system is one of the effective tools to alleviate this problem.It captures users’ interests and preferences through the interaction sequence between users and items,and recommends appropriate items for them.However,due to the anonymity and lack of information in session sequence data,how to make full use of limited session sequence data to improve the accuracy of recommendation has become an important research task of session sequence recommendation.Aiming at the problems of the high computational complexity of self-attention and unable to accurately measure the importance of items in the existing session-based recommendation methods combined with self-attention,a multi-interest aware adaptive self-attention network model is proposed for session-based recommendation.In the model,an adaptive self-attention network is designed to improve the sparsity of the output item importance score by introducing a class of highly sparse computing functions to replace the softmax function.At the same time,considering the differences of each sequence,a sequence adaptive factor module is designed to enable each sequence to dynamically select the appropriate computing function based on the context,thus enhancing the ability of the model to accurately measure the importance of items.To solve the problem of the high computational complexity of self-attention,an interest aggregation layer is proposed to capture constant user’s interests,which reduces the computational complexity by converting the correlation between items into the correlation between items and constant interests.In addition,the model adopts a decoupled position coding scheme to reduce the coding coupling between items and position information,to avoid the introduction of noise information.Finally,to fully consider the user characteristics,the user global preference module is designed to capture the user preferences.Experimental results on several public datasets show the effectiveness and feasibility of the proposed model.Compared to the baseline methods,the maximum gain percentage of the proposed model is 6.54%on HR@10 and 5.71%on NDCG@10,respectively.Aiming at the problems that the session-based recommendation method combined with graph neural network in e-commerce scene adopts the construction method of session graph with information loss,and fails to fully consider the influencing factors of feature modeling,a price-aware information lossless model is proposed for session-based recommendation.The model adopts a method of constructing a session graph with information lossless to fully encode the session information and reduce the information loss during construction.To explore the new factors that affect users’ preferences,a price tolerance factor module is designed to model users’ price tolerance for various items.In addition,a new user intention coding scheme is proposed to capture user intention from the item category level to improve the accuracy of capturing user intention.Finally,to fully consider the modeling of user characteristics,a dynamic interest module is designed to capture the dynamic interest of users over time.Experimental results show that the proposed model achieves better performance compared with multiple representative baselines.The maximum gain percentage of the proposed model is 1.52%on HR@20 and 1.58%on MRR@20,respectively.