Matrix Factorization Recommendation Algorithm Based On Subject Feature

The main significance of personalized recommendation is to help people mine the information that users may be interested in from the massive information data,save the time of searching information,improve the efficiency of users,and improve the user experience.The subject of a recommendation system mainly includes users and items.In the process of personalized recommendation,it is often necessary to compare the similarity between subjects,so as to provide basis for personalized recommendation.However,the traditional collaborative filtering algorithm only considers the common score items between users,neglects the relationship between users and items,which makes the recommendation result not ideal.The relationship between the subjects of the recommendation system can reflect their own characteristics.On the one hand,the user's historical rating of the items can well quantify the user's preference characteristics.On the other hand,the user's tags for the items also provide the real basis for the construction of the characteristics of the items.Therefore,based on the interaction between subjects to improve their own characteristics can better improve the accuracy of recommendation.Based on this idea,this paper proposes a matrix decomposition recommendation algorithm based on the subject feature.The main work is as follows:1.Aiming at the problem that the traditional funk SVD model doesn't consider the specific details of the subject and the interaction records between the subjects,a matrix factorization model combining the user preferences is proposed.The model first constructs the user preferences based on the historical interaction records of the user and the items,and then combines the feature differences of the subject with the matrix decomposition model to train the user characteristics and the items characteristics.When constructing similar user sets,the user features obtained from matrix factorization are combined with user preference features constructed from historical interaction records and user rating confidence of different users to improve the accuracy of recommendation.2.In this thesis,we propose a method to construct the features of items based on the confidence of users' tags,which takes the tags as another form of the features of items.Combined with the user-tag confidence of different users,the user's heat is distributed to the items through the historical tag records,and the mass diffusion method and heat transfer method are used for heat distribution.The user-tag confidence is used to measure the user's ability to use the tags,and as the initial heat of the user node.Finally,according to the amount of heat allocated from the tags node,the tag features of the items are constructed.3.According to the user rating and the characteristics of the item tag,the user tag feature is constructed,and combined with the user feature and the item feature,an improved matrix factorization recommendation algorithm is proposed.And the algorithm is tested on the movielens data set,and compared with SVD + + and other algorithms in the same data set.The experimental results show that the matrix factorization recommendation algorithm based on subject feature has improved in the evaluation indexes of RMSE and precision.