Research On Recommendation System Algorithm Based On Hybrid Similarity Measurement And Attention Encoding Mechanism

The recommendation system is a system widely used in the current Internet information industry.Facing a variety of recommendation system application scenarios,data characteristics and startup status,the construction of collaborative filtering recommendation system faces a series of serious problems,including but not limited to cold start problems,sparse user / item matrix and dimensional explosion problems,User-item organization long-tail distribution problem,user-selected similarity measurement problem,etc.Aiming at the problems of the current collaborative filtering recommendation algorithm,by introducing a hybrid similarity measurement method,combined with the concentration encoding mechanism to encode user behavior data and other methods,the cold start and user / item matrix that are common to traditional recommendation algorithms are alleviated to some extent Sparse problem.The recommendation algorithm ACFSI based on mixed similarity metric and concentration neural network coding proposed in this paper shows more efficient and accurate recommendation effect on the Movie Lens 1M data set compared with the traditional recommendation algorithm.The work of this article mainly includes:(1)A weighted similarity calculation method(CFSI)based on hybrid similarity calculation is proposed.This algorithm improves the accuracy of the user-oriented and item-based similarity calculation during the collaborative filtering algorithm calculation process.A traditional similarity measurement method,which summarizes and summarizes the emphasis of various similarity measurement methods on user feature comparison,and finds a hybrid similarity measurement method that takes into account multiple expression angles of features,which can reasonably take into account the unique characteristics and Type items share characteristics and effectively avoid noise.At the same time,through this algorithm,the effect of cold start on the accuracy of the recommendation system is alleviated,and the similarity measurement level of the recommendation algorithm performance is improved.(2)An encoder(ARSE)based on the concentration mechanism and recurrent neural network is proposed.The encoder can compress the matrix according to the inputfeature matrix.A feature encoder is used to replace the user in the traditional collaborative filtering system.The item similarity matrix itself encodes the variable-length user item rating sequence into a fixed-length user feature expression vector,and replaces the huge sparse matrix with each feature expression vector combination,so as to achieve the effect of reducing the resource consumption of the collaborative filtering algorithm.At the same time,by adding the user / item's own characteristics to the encoder input,the adverse effects of "loyal users" and "hot items" in the traditional collaborative filtering algorithm are alleviated to a certain extent.(3)According to the CFSI and ARSE algorithms proposed in this paper,a new recommendation algorithm(ACFSI)with a cascade structure is constructed.This algorithm first accepts the data of the user's rating of the item;through the rating data,it constructs and trains to output fixed-length user features Vector-targeted neural network encoder(ARSE);after that,the encoder output user feature expression will be imported into the hybrid similarity measurement algorithm,through the hybrid similarity-based measurement algorithm(CFSI).The collaborative recommendation framework compares the feature expression vectors of each user to obtain a sequence of recommended items for each user.Complete the recommended task.The cascade algorithm improves the recommendation accuracy and at the same time compresses the space complexity of the algorithm,which provides a basis and reference for the research of recommendation algorithms and related similar problems.