Research On Recommender System Fused With Item Label

With the change of the times,the development of technology is very rapid,a variety of resources and items overwhelm the user,the recommender system was born at this time as the user's right-hand man.The recommender system is able to extract some potential characteristics based on the user's historical behavior,and recommend some resources and items that meet the user's needs and personality preferences to improve the recommendation accuracy.However,due to the huge amount of information on the web,users have access to only a small portion of the resources and can only rate or evaluate very few of them.This results in extremely little rating information available to the recommender system,which can cause difficulties such as data sparsity and cold start.To alleviate these problems,there is a growing number of websites that use social tagging to improve the quality of recommendations,and users are free to add their own tags to the resources they have access to and prefer.By introducing tagging information into the recommender system,it is possible to quickly and accurately find the set of users whose preferences are more similar to those of the target user and the set of resources that fit the user's personality,thus filling the scoring matrix,alleviating the problem of sparse data in the recommender system,improving the interpretability of recommender and gaining user trust.Tags can cover the characteristics of resources and provide a reliable basis for classifying resources,and by using tags,users can easily retrieve a collection of resources that match the tags.Traditional information retrieval on the web is done by searching for keywords,but this does not reflect the user's thoughts.By searching through tags,you can better match your ideas and opinions and make your search more accurate.The non-linear matrix factorization model is a classical deep learning recommender algorithm that combines the traditional matrix factorization GMF and the multi-layer perceptron MLP to facilitate learning features in various dimensions.The matrix factorization algorithm is used to obtain predicted scores for unrated resources by decomposing the scoring matrix into a matrix of user features and resource features and continuously iterating.However,the above recommender algorithms do not consider the role of label information,the data sparsity problem is serious,and the recommender accuracy is not high.To address the above problems,the following research is conducted in this paper.(1)A nonlinear matrix factorization optimization model NLTMF(Nonlinear Matrix Factorization Model for Integrating Item Tag)incorporating item label content is proposed.The method incorporates item label content in the nonlinear matrix factorization,and uses Autoencoder self-encoder to downscale the item label data and extract features,while incorporating Batch Norm in the embedding layer,which makes the model faster to train,more interpretable,and more accurate.stronger interpretability and higher recommender accuracy.The experiments on three datasets,movielens-100 k,movielens-1m,and mllatest-small,show that the algorithm has good prediction effect.(2)A matrix factorization algorithm model UTag JMF(Matrix Factorization Algorithm Based on The Similarity of User Tag Sets)based on user label set similarity is proposed,which firstly combines the rating information and label information to construct a user label set similarity matrix,and then calculates the user similarity by the Jaccard algorithm,and finally combines the two and performs matrix factorization at the same time.Experiments on the ml-latest-small dataset and on the hetrec2011-movielens-2k dataset show that the algorithm model has significantly better predictions than other algorithms in terms of root mean square error and mean absolute error.(3)A recommender system platform application based on the similarity of user tag sets UTRS(User Tag Similarity Recommender System)was designed to recommend similar users and their possible interested resources to the target users through the similarity between user tag sets,and to recommend the resources with the top prediction scores to the users,thus realising personalised recommender and improving the interpretability and accuracy of the recommender.