A Collaborative Filtering Recommendation Algorithm Based On Implicit Feedback Data

With the rapid development of Internet technology,people can quickly obtain a lot of information through the Internet.In order to avoid the problem of "information overload",recommendation system technology has been widely studied and applied.At present,recommendation system has become one of the most popular research fields in machine learning.However,the existing collaborative filtering recommendation algorithm fails to deal with the deviation of rating prediction caused by MNAR(missing not at random).Meanwhile,the popular deep learning recommendation algorithm is limited by the computational complexity and the acquisition of additional information,so it can not be applied on a large scale.At this time,studying the algorithm of recommendation system and proposing an algorithm which can effectively deal with the problem of MNAR will promote the development of existing algorithms and bring better recommendation system.For the problems of existing research,we improved popular rSVD(Regularized Singular Value Decomposition)and gSVD(Group-special Singular Value Decomposition)based on the idea of multi-task learning(MTL),then proposed mSVD and mgSVD.And the specific optimization methods were given.These two methods effectively integrate the information of the implicit feedback data,and can better deal with the problem of non-random missing,so that they can achieve better results.We conducted a large number of simulation experiments on multiple sets of simulated data,which showed that the mSVD method and mgSVD method can effectively use the information of implicit feedback data,and achieve significantly better results than the rSVD method and gSVD method.And the advantages of these two methods are more obvious when the missing rate of rating is higher.Finally,through the analysis and experiments on real-world recommendation system data,we showed that the problem of missing not at random is widespread in real-world recommendation systems,and these two methods in this paper are also significantly better than existing methods in real-world data.