Research On Recommendation Algorithm Based On Category Preference And Representation Learning

The recommendation system brings convenience to people in the era of information overload,that makes the recommendation algorithm become a research hotspot.The collaborative filtering recommendation algorithm has been widely studied for its simplicity and convenience.However,the traditional collaborative filtering recommendation algorithm has data sparsity problems.For this problem,the usual practice is to add the auxiliary information of users and items.The category preference and information in the heterogeneous networks of users and items are used to improve recommendation performance.The following two recommended algorithms are proposed:The traditional category information-based recommendation algorithms often use the category information as auxiliary information,and do not consider the many-to-many relationship between items and categories.Aiming at this problem,a matrix factorization recommendation algorithm based on combinational category space is proposed.First,a combinational category is defined to make the items and the combinational categories have a one-to-one relationship.Secondly,the semantic relationships and distances are defined in the combinational category,and the user's browsing behavior,the categories of the items,and the relationships between the users and the items are better integrated.Finally,the implicit feature model of users and items based on matrix factorization is established to better represent the relationship between combinational category and user interest preferences.The experimental results show that the recommended performance is significantly improved.In order to solve the problem that heterogeneous information is difficult to extract in the recommendation algorithm based on heterogeneous network,beside,use neural network to learn the interaction information between users and items,a deep learning recommendation algorithm based on heterogeneous network feature representation is proposed.First,heterogeneous information networks of users and items is built.Interactive meta-paths of users and items are defined,that make heterogeneous information easy to extract,and better align user browsing behavior with various information.Then,the feature representation of the heterogeneous network is used as the input of neural network.The LSTM and Attention mechanisms are used to simulate the recent interest changes of users and the popularity of recent items.Experimental results show that the recommended performance is improved.