The Research And Implementation Of Recommendation Algorithm Based On Multi-tag And Classification Sorting

Recommender systems,one of the most popular research fields in recent years,involve different fields such as information retrieval,data mining,machine learning,complex network and sociology.Although recommender systems have been developed by leaps and bounds,it still faces problems like cold start,sparseness,scalability,etc.As the data of recommender systems become richer and richer,recommender systems incorporate a variety of data information into recommendation algorithms in order to get better recommendation.Tag information is one of the most important information.This paper uses Tag information as a point for research.In score forecasting scene,this paper proposes an approach that incorporating the multidimensional tag information into neighborhood-based collaborative filtering and matrix factorization to improve algorithm's ability to predict score and alleviate the problems of cold start and sparseness.This approach combines tags according to the relation between tags,and combines user tags and item tags information to build multidimensional tag information.For neighborhood-based collaborative filtering,this paper uses the user and multi-dimension tag matrix to calculate user similarity,and uses the item and multi-dimension tag matrix to calculate similarity of the items.And then we mix collaborative filtering with the similarity of the multidimensional tag by linear weighting method and predict score.Collaborative filtering based on matrix factorization is a popular recommendation algorithm.The effectiveness of matrix factorization can be enhanced by implicit feedbacks.This paper takes multidimensional tag information as an implicit feedback as well.User and multidimensional tag matrix is converted into Boolean Matrix and then be incorporated into matrix factorization model.In Top-N recommendation scene,recommender systems give users a list of items which users might be interested in,it is closer to the nature of recommender systems.This paper proposes a classification sorting approach to Top-N recommendation based on the idea of sorting learning.It uses a hybrid recommendation named waterfall framework,then employs logistic regression to sort recommendation list based on the result of Collaborative filtering.Logistic regression takes tags and tag combinations as input features,and classify whether users are interested in items,and then sorts recommendation list by output score.Moreover,this paper uses popular online algorithm FTRL(Follow the Regularized Leader)to optimize logistic regression which classify and sorting based on collaborative filtering result.Experiment results show that the collaborative filtering based on multidimensional tag we proposed improves score prediction effectiveness of recommendation algorithms,and partly alleviate the problems of cold start and sparseness.And the classification sorting approach we proposed for Top-N recommendation can improve the effectiveness of recommendation list.