Research On Deep Matrix Factorization Method Based On Missing Not At Random Data In Social Networks

In the information age,the use of search engine alone does not meet the needs of our daily life any more.Recommender systems,brought about in the context of Internet,solve the above problem well.They volunteer the information that people may be interested in by studying their historical behaviors.However,common recommerdation algorithms can not deal with coldstart problem and heavily suffer from data sparse problem.With the rise of online social networks,researchers start to incorporate the influence of social networks into recommerdation algorithms to cover the shortage of common methods,which is called social recommendation.This paper is mainly divided into two parts: matrix completion and deep social matrix factorization.Both of them focus on digging deeper in detail the features of users and items through studying the trust relationship between users and interactions between users and items.Common recommendation methods always assume that the data is missing at random,which leads to deviation.Matrix completion aims at completing a part of the missing data in the original datasets to avoid the drawback.After matrix completion step,considering the importance of taking into account both explicit ratings and implicit feedback,this work proposes a framework that extend Deep Matrix Factorization(DMF)model to social recommendation.With the rapid development of online social networks,a mass of data mining techniques for integrating social relations into ordinary recommendation algorithms have emerged.Despite the extensive studies,social recommendation methods always assume that the data is missing at random(MAR),which is in the rare instance.When the assumption is invalid,it will definitely damages the recommendation effectiveness.Taking the affinity relationship between members of the network as the starting point,this work proposes a novel matrix completion method based on Baysian Personalized Ranking,aiming at completing a part of the missing data of the original dataset.The results of the experiment show that data pre-processing can further enhance the performance.Especially for Ciao dataset,the MRR value can be increased by 6.36%.Recommender system datasets can be separated into explicit ratings and implicit feedback.Compared with implicit feedback,explicit ratings are in less quantity,and only using explicit ratings to recommend is more effective.Comparing with explicit ratings,implicit feedback is talented about ferreting out users' hidden interests.But only a little of the existing work has taken into account both explicit ratings and non-preference implicit feedback when it comes to social recommendation.In the view of the advantage of DMF model in common recommender systems,we extend the DMF model to incorporate the user-user trust relationship and fully exploit the explicit ratings and implicit feedback.Extensive experiments conducted on three real-world datasets demonstrate that our proposed method provides the best top-N recommendations,illustrating the benefits of the added model complexity.