Design Of Recommendation Algorithm Based On Multi-feature Implicit Feedback

With the rapid development of Internet technology,the amount of data in the network is increasing exponentially.In the case of information overload,it is difficult for users to find interested items in a timely and accurate manner.Recommender systems can filter a large amount of irrelevant information for users by analyzing the relevant characteristics of users and items,and recommend items that meet their preferences.The accuracy of recommender systems is often limited by two major problems.One is data sparsity caused by that the amount of rating is too small,and the other is cold start caused by new items or new users.This paper mainly researches into recommender systems for these problems to improve the accuracy of recommendation results.This paper is based on the multiple types of implicit features related to users and items in recommender systems.On the one hand,this paper deeply analyzes the implicit relationships in ratings to make full use of ratings.On the other hand,it combines transfer learning to study relevant implicit feedbacks in auxiliary domains and use them to assist recommendation of the target domain.From the perspectives of in-depth mining of ratings and introduction of auxiliary information,two innovative recommendation algorithms are proposed.The specific research work and innovations of this paper are as follows:(1)This paper analyzes implicit relationship between users and items,designs and implements a recommendation algorithm based on the implicit relationship between users and items.A user's rating of an item can not only clearly express how much he likes the item,but also implicitly reflect the relationship between him and other similar items.The paper studies a multiple layer method of defining the implicit relationship between users and items,and uses the implicit relationship as the criterion of item classification in the pairwise algorithm,so as to find more implicit information in ratings.Then,the classified item sets are combined with the pairwise algorithm,and the two scales of set and item are used as the unit for pairing comparison.Results of experiments based on three real data sets show that this algorithm can provide more accurate recommendation results compared with other algorithms.(2)This paper analyzes the relationship between ratings and other heterogeneous data,and designs a cyclic transfer learning recommendation algorithm framework based on heterogeneous feedbacks.Users will leave a large amount of historical data when using recommender systems.In addition to ratings,there are often other types of data,including user's attitude information about items,trust relationships between users,and so on.The paper takes rating as the target domain and other heterogeneous data as auxiliary domain,extracts features in the auxiliary domain that are helpful to rating prediction in the target domain,and innovatively designs a cyclic two-way transfer learning method to promote the full exchange of knowledge between the auxiliary domain and the target domain.Based on the real situation,this paper implements two specific recommendation algorithms in combination with the attitude auxiliary domain and the trust relationship auxiliary domain.Results of experiments based on two real data sets with trust relationship show that the recommendation performance of two specific recommendation algorithms can achieve a significant improvement compared with other algorithms,and can overcome cold-start problem.This paper uses 26 figures,11 tables and 60 references.