Research On Collaborative Filtering Recommendation Algorithm Based On Item Semantics

In recent years,the application of recommendation systems on the Internet has become more and more widespread,and has also received more attention from many experts and scholars.The recommendation system can help users quickly find information that meets the individual needs of users from massive data.However,with the advent of the era of big data,existing recommendation algorithms face many new challenges,such as data sparsity and scalability problems caused by the huge increase in the number of users and projects;new users or projects have just joined the recommendation system A cold start problem occurred in.In response to these challenges,this paper conducts research,and its main work is as follows:(1)A item semantic representation model based on improved TF-IDF is proposed.For the TF-IDF algorithm that only counts the frequency of feature words,without considering the shortcomings of the meaning of the feature words themselves,this paper uses the semantic similarity between feature words to construct a method to calculate the frequency of semantic words to improve the TF-IDF algorithm,and further strengthens the The item description text has important feature word weights,and then combined with the word2 vec model weighted average to calculate the semantic vector of the entire item,making the item's semantic features more distinctive and convenient to distinguish the differences of items semantically.(2)A matrix factorization algorithm based on item semantic representation model is proposed.Based on the matrix factorization,additional auxiliary data item description text information is introduced.The item semantic vector is calculated by the item semantic representation model,and then a matrix of user semantic preference potential features is modeled in the matrix factorization to achieve item semantics is integrated into matrix factorization,and subjective influencing factors on user semantic preferences are combined in predictive scoring.Experiments show that the algorithm improves the accuracy of prediction rating and effectively alleviates the impact of data sparsity.(3)A cold-start recommendation algorithm based on clustering segmentation of completion rating matrix is proposed.Based on the matrix factorization andcompletion rating matrix based on the item's semantic representation model,for the cold start problem,the cold start recommendation is collaboratively filtered using the information of users and the inherent characteristics of the item.Use user attribute information and item semantic information to construct user and item similarity calculation methods,and use the clustering algorithm k-means to cluster users and items respectively.According to the clustering results,divide the user-item rating matrix and divide the data.The original user-item matrix with high dimensions is divided into multiple small matrixes with lower dimensions.Finally,the data in the small matrix is ??used for collaborative filtering and cold start recommendation.This algorithm not only solves the cold start problem,but also improves scalability and real-time performance.