Cross-domain Recommendation With Tag Transferring

In the era of mobile Internet,the channels for users to obtain information are increasing day by day,so it is difficult to make practical judgments on content quickly.The emergence of problems such as information overload has led to the development of the recommendation system and has been widely used in various domains.However,issues such as data sparsity and cold-start of the recommendation system are still difficult to be effectively solved.As a result,more and more machine learning techniques are beginning to be used to improve the result of the recommendation algorithms.Transfer learning can assist the target domain to complete the recommendation task by transferring knowledge from other domains,which can increase the learnable data,strengthen the recommendation model for the recommendation results,and alleviate the data sparsity and cold-start problem in the target domain.Therefore,this paper develops a cross-domain recommendation model and algorithm based on knowledge transfer.The main works are as follows:(1)Most cross-domain recommendation algorithms often use transfer learning technology.The existing transfer learning methods are mostly based on a single rating model.In the case of unrelated domains,it often leads to negative transfer and poor recommendation results.This paper combines rating knowledge with behavioral knowledge to assist in learning tasks in the target domain.In behavioral data,the item tag is related to the user's real preferences and can reflect the implicit characteristics of the user or item from another perspective.In order to alleviate the negative transfer,we aggregate the tags of the auxiliary and target domains to obtain the transferable knowledge.Based on this,an Item-based Tag Transfer Collaborative Filtering algorithm is proposed.The algorithm discards the single auxiliary method of transferring only the rating mode,combines the two kinds of heterogeneous knowledge of user rating and item tag.Incorporating the tags of item alleviated the data sparsity of the target domain to a certain extent and also solved the negative transfer problem.(2)In the real world,users are more concerned with the results of recommendations than ratings,and the recommendation by the predicted rating does not accurately capture the user's preference.Based on this,a knowledge transfer recommendation algorithm RBT for list-wise ranking is proposed.The recommended target of the algorithm is shifted from rating prediction to ranking prediction.By transferring domain information including ratings and tags can get more item preferences,which can build richer partial-order item pairs to train the ranking model and ultimately generate a candidate recommendation list for the target user.(3)The data set of Douban Movie,Douban Book,MovieLens and other data sets were selected to carry out the comparative experiments of the above algorithms.The experimental results show that ITTCF has improved RMSE by 1.61%-6.67%and MAE by 1.97%-8.83%comparing to the baseline model.Through learning to rank,RBT has also been improved in accuracy,recall,NDCG and other ranking-oriented recommendation evaluation indicators.