Research On Long-tailed Items Recommendation Based On Real-valued RBM

The recommendation system is an effective method to solve the problem of information overload.The core is to use the recommendation algorithm to provide users with personalized services.Commonly used recommendation algorithms include collaborative filtering,content-based recommendation algorithm,latent factor model,and hybrid recommendation.However,most of the recommended algorithms tend to be recommended for popular items,so that the recommendation of long-tailed items still faces serious challenges.For the recommendation of long-tailed items,this paper studies this issue from two perspectives.Firstly,from the perspective of the relationship between the users and the items,the implicit relationship between the users and the items is used to explain the relationship between the users and the items.Secondly,from the perspective of the relationship between the users,the user's preference is the basis of the topic.Using collaborative filtering based on improved real-valued RBM,the relationship between users is constructed to realize the prediction of unknown topics of the target users.The specific research content is as follows:(1)Through the analysis of the characteristics of long-tailed items,a three-tier recommendation model of "users-topics-items" based on long-tailed itmes recommendation is proposed,and the recommended model architecture is divided into two parts: "users-topics" and "items-topics".The "users-topics" section uses the topic model to extract themes based on the user's historical behavioral data to construct a user interest preference theme model.The "items-subjects" section uses the subject model to extract themes and constructs the items based on the item content data.Feature topics,finally calculating the similarity between user interest preference topics and items feature topics,to achieve recommendations for long-tailed items;(2)In order to overcome the topic model's recommendation in long-tailed articles,it does not have the disadvantage of guiding users to discover new areas of interest.Based on user preference topics,a collaborative filtering approach based on improved real-valued RBM is proposed to build a user-preference topic prediction model.The target user does not know the prediction of the preferred topic to discover the user's new interest area;(3)By analyzing the characteristics of user preference topics,it is known that because the same user corresponds to a number of different topics,there may be the same topic words between different topics.The neurons of the traditional real-valued RBM visual layer cannot correspond to the topic words one by one,and each real-valued RBM corresponds to a topic of the user,and the same topic words share the weight and bias.Set.In addition,the visual layer of the traditional real-valued RBM is sampled from the Gauss distribution to fit the distribution of the training samples.Because the key words extracted in this paper conform to "word distribution"(polynomial distribution).The real-valued RBM can fit the distribution of the training samples directly,and the activation function of the ReLU as the real-valued RBM visual layer is put forward,and the experimental knot is put forward.The results show that the improved real-valued RBM has better prediction ability for user preference topics.(4)When using real-valued RBM to predict topics,the topic is the prediction target,each user's preference corresponds to multiple topics,and each topic corresponds to a real-valued RBM,and each user is assigned a real-valued RBMs.By using the average real-valued RBM weight value of the same topic word in the neighboring user as the real-valued RBM weight of the topic word in the topic to be predicted topic.