Research And Application Of Matrix Factorization In Recommender Systems

Since mankind entered into the information era in the late twentieth Century,we've witnessed flourishing and prosperity as never before in Internet development.Human daily activities become more closely linked with the Internet.For one thing,people enjoy the convenience the Internet brings,for another,the data on the Internet is getting more and more,leading to difficulty in finding the information they need in the vast amounts of data,which is the so-called "information overload" problem.With the birth of personalized recommendation system,the above problems have been effectively alleviated.The most common algorithm is CF algorithm,which utilizes the item scores rated by numerous users to generate recommendations.Since the recommendation algorithm based on Matrix Factorization was proposed in 2007,researchers have devoted a lot of energy into the MF model,which has sprung up in the field of recommendation.The most obvious merit of Matrix Factorization model is that it can maintain better recommendation performance while the datasets are relatively sparse.However,the recommendation results are tough to explain,and sometimes it is difficult to convince users.In chapter3,four common Matrix Factorization models are introduced and analyzed in detail,their recommendation performance is quantitatively compared on the Movielens dataset.The final results prove that Probabilistic Matrix Factorization outperforms the other three methods on Prediction and Ranking task.When exploiting implicit feedback in recommendation system,there are no abundant negative samples in the training set,which results in the significant over-fitting problem,but if considering all the items as negative samples,the results will suffer a large bias,which is called “one class collaborative filtering” under this circumstance.In this thesis,a sampling method(NegSec algorithm)leveraging popularity,social networking and user reviews is proposed,to distinguish the items users dislike(negative samples)from the items which users have no behaviors by using these auxiliary information as much as possible.The algorithm is tested on the Douban dataset,and the MF model,trained by the negative samples selected by NegSec,can achieve 15.7% in terms of Precision,which is twice as the random selection algorithm does.This thesis has also done some work in combining matrix factorization with deep learning,proposes a joint review modeling by deep learning and matrix factorization recommendation algorithm called DLPMF.Convolutional neural network is applied to extract the features of the user reviews,and DLPMF links the extracted features with the item latent vector in Matrix Factorization,whose essence is to add a new complex regularization term in the original cost function.A number of experiments on douban dataset demonstrate that DLPMF effectively utilizes the user reviews,and achieves 0.9 on RMSE,which is superior to PMF in score prediction.