Research On Personalized Recommendation Algorithm Based On Matrix Completion And Graph Embedding

Personalized recommendations can provide contents that users may be interested in based on their hobbies,which is a technique for solving the problem of information overload.Collaborative filtering is an important driving strategy in the recommendation system.Collaborative filtering algorithm calculates user/item similarity based on the user-item rating matrix,and sets the users/items with high similarity as the neighbor of the target user/item,and then recommends the items that have not been rated to the target user according to the ratings of the neighbors.However,the collaborative filtering algorithm is built on a user-item rating matrix,and user/item similarity only depends on the user's ratings for the items.Since collaborative filtering relies on the user-item rating matrix too much,most users in the recommendation system only evaluate a few items in real life,and most of items are only evaluated by a few users,which leads to the user-item rating matrix with a large sparsity,making it difficult to obtain a high recommendation quality.In addition,when a new user enters the system without rating record,the collaborative filtering algorithm cannot recommend.This is called cold start.Therefore,it is a problem that how to recommend new users in the case of cold start to be solved.In order to solve the problem of rating matrix sparsity of collaborative filtering algorithm,this paper utilizes LMaFit algorithm for matrix completion,reducing the sparsity of the matrix and alleviating the problem of data sparsity;and calculates users' similarity based on their attributes and ratings,which enables recommender system to make recommendation for new users,thus alleviating the phenomenon of cold start.Based on the above research,k-means++ algorithm is used to cluster the users,finding the neighbors of the target user,predicting the ratings of users.The experiments based on Movielens-100 k datasets show the proposed algorithm has good performances in MAE and RMSE.In order to learn the deeper latent representation of users and items,and understand the nonlinear structural features between users and items,this paper introduces the graph embedding to the traditional collaborative filtering algorithm,and learns the embedding representation of users and items based on DNGR model.And this paper calculates the relevance of user features and item features.The experiments based on Movielens-100 k datasets show the quality of recommendationof proposed algorithm is better than the traditional collaborative filtering algorithm.