Research On The Convolutional And Self-Attention On Session Recommendation

In the background of big data,recommendation system is one of the key technologies to solve the information overload problem and facing the situation that user identity is not available,the traditional recommendation based on user identification or rating records will no longer be applicable,and session-based recommendation based on user behavior is an effective solution.In session-based recommendation,the input of the model is mainly the user’s behavioral information,which can be divided into two kinds according to the time span,one is the specific interest in the short-term time,and the other is the interest preference represented by the long-term behavior.It is very challenging to utilize the short-term and long-term behaviors of users in a better way to express the complete behavioral intention of users,which can improve the effectiveness of recommendations.Faced with the situation that user identity is not available,the traditional recommendation based on user identification or rating records will no longer be applicable,and the session recommendation based on user behavior is an effective solution.In the session-based recommendation,the input of the model is mainly the user’s behavioral information,which can be divided into two types according to the time span,one is the short-term interest within a specific time,and the other is the interest preference represented by the long-term behavior.It is a very challenging problem to better represent the short-term and long-term behaviors of users and express the complete behavioral intention of users,so as to improve the effectiveness of recommendations.Existing methods for long-term user behavior mainly use recurrent neural networks or attention networks to model users’long-term behavioral preferences.Recurrent neural network is difficult to learn item dependencies from a long distance sequence,and traditional self-attention networks have the problem of long-tail distribution of attention weights in the process of calculation.Based on the above problems,this paper proposes a Probabilistic Screening Self Attention Network(SSA)to model users’long-term historical behaviors.Which screens out unimportant weight associations and need fewer computational and memory resources.The time complexity of the traditional self-attention network is reduced from(~2)to(7)7)7)7)7)7)),and the interference of weight noise to the network is reduced,while the effectiveness of the improved modeling method is verified through experiments.Session recommendation is a problem of sequential representation and interaction between short-term and long-term behaviors of users,so further research and improvement of long-term historical behaviors and short-term behavioral features representation of users are needed to better express the complete behavioral intention of users and thus improve the effectiveness of recommendations.Accordingly,this paper introduces a compound convolutional scheme for modeling short-term behavioral information in sessions,and on this basis,we combine the previously proposed improved long-term behavioral modeling approach(SSA)to construct a network architecture based on Compound Convolutional and Probabilistic Screening Self Attention Network(CCNN-SSA).The architecture utilizes convolutional modules with different spatial dimensions to extract complex short-term behavioral features,while probabilistic multi-headed self-attention is used to learn long-term behavioral interactions from the sessions.In order to verify the proposed model,the proposed model is validated on two public e-commerce benchmark datasets.The experimental results show that the effectiveness and rationality of the proposed CCNN-SSA model are verified through experiments.