Research On The Recommendation Method Of Using Side Information And Unlabeled Data

With the popularity of the Internet,especially the ubiquitous interconnection between emerging mobile applications,various types of data are showing explosive growth,and it is becoming more and more difficult to quickly obtain useful information from them.In this context,the recommendation system came into being and played an important role in combating information overload.Collaborative Filtering is one of the key technologies for constructing a recommendation system.Its core idea is to mine the user's potential information preferences from the historical interaction data(such as ratings)of “user-item”,and predict the future possibility for the user accordingly Items of interest.However,compared with the huge user and item scale,the observable rating data is very sparse,which seriously restricts the recommendation performance of the collaborative filtering algorithm.In order to cope with the problem of data sparseness,scholars have proposed to use edge information or unlabeled data to improve the recommended performance of collaborative filtering algorithms.Both methods have their own advantages,but there are few mixed research schemes combining the two ideas.In view of this,this thesis proposes a collaborative filtering scheme that combines side information with unlabeled data to deal with data sparseness in a more effective manner,thereby obtaining a greater degree of recommendation performance improvement.In the collaborative training framework,the scheme uses a collaborative filtering method based on user neighbors and a collaborative filtering method based on item neighbors to initialize two base recommenders;in subsequent iterations,each recommender independently predicts unlabeled data and believes the prediction Several pseudo-label samples with higher degrees are added to the other party's training set and retrained to obtain a progressive improvement in recommended performance;the process is repeated until convergence.At the same time,the solution also incorporates social edge information in the user's neighbor method,including using social information to improve similarity calculation,update neighbors,and upgrade prediction formulas,etc.,in order to increase the difference between the two base recommenders,thereby achieving more Good collaborative training effect.In addition,this thesis also adds a confidence verification link for pseudo-labeled samples in the collaborative training framework to prevent the performance degradation of the semi-supervised learning algorithm due to the misuse of noisy samples.Experimental results show that the combined use of side information and unlabeled data can effectively alleviate the problem of data sparsity.The recommended performance of the scheme described in this thesis is significantly better than other collaborative filtering techniques that use side information or unlabeled data alone.The main contributions of this thesis can be summarized as the following three points: 1)verify the possibility and effectiveness of the joint use of side information and unlabeled data,and provide new ideas for alleviating the problem of data sparseness;2)propose a joint use under the collaborative training framework The collaborative filtering scheme of unlabeled data and social side information greatly improves the performance of the recommendation system;3)Combining semi-supervised learning with a variety of social network analysis technologies,to a certain extent,enriches and improves the research content in the field of social recommendation systems.