Research On Session-Based Recommendation Model Integrated With Multi-Feature Information

With the explosive growth of Internet information resources,the problem of "information overload" has become increasingly prominent in many Web applications such as search,ecommerce and video websites.As an effective way to solve the problem of "information overload",recommender system has become the focus of attention in academia and industry and has been widely used.At present,most of the recommendation systems are based on the user's personal information and historical behavior data.However,there are many scenarios in which the user's identity may be unknown and only user behavior data in the current session is available.The session-based recommendation is used to deal with scenarios where the user's identity is unknown.Most of the traditional recommendation methods based on sessions do not consider the click order of items in the session sequence.However,although the deep learning method based on session effectively models the click sequence information of the session,it often ignores other information of the item,such as the category information.Most of the session-based recommendation methods do not treat the click sequence of user sessions as a directed graph,so they cannot obtain the relationship between nodes,and do not consider the interest offset during the session.To solve the above problems,this paper puts forward a novel session-based recommendation model.The main work and innovation points are as follows:(1)A session-based graph neural network recommendation model is proposed,which constructs the user's click sequence into a directed user session graph.The structure and information of the user session graph were learned through the Graph Neural Network(GNN).A Gated Recurrent Unit(GRU)was used to extract the timing information in the click sequence and the interest information during the entire session.Finally,the interest information at the last moment in the user session were combined to make recommendations for the user.(2)A session-based recommendation model that integrates multi-feature information is proposed.The directed item session graph and the corresponding directed item category session graph of the user's session click sequence are constructed.The attention network is used to fuse the embedded vector of the obtained items and item categories.Using the fused vector for subsequent recommendations.By adding item category information as a supplement to item information,the impact of user interest deviation is reduced,and the effect of the model is further improved.(3)Experiments were carried out on the Yoochoose 1/64 datasets,Yoochoose 1/4 datasets and Diginetica datasets.Firstly,the two models proposed in this paper are compared with each other to verify the effectiveness of adding interest information and category information.Then experiments were carried out with eight benchmark models to verify the improvement of the proposed model in the recommendation effect.Both the Precision and MRR evaluation indicators proved the effectiveness and accuracy of the model proposed in this paper are discussed.