Research On Matrix Factorization Recommendation Algorithm Based On Implicit Feedback And Temporal Information

With the development of the Internet,E-commerce has brought a lot of life convenience to users,while some problems hinder them from obtaining the necessary products,such as tons of data,information multiplicity,and homogeneity.The above phenomenon is called information overload,and recommender system can solve this issue efficiently by analyzing the past behaviors and predict the products the users have not bought but suit their interests.Past behaviors are composed of explicit feedback and implicit feedback.Explicit feedback is a series of metrics that can directly reflect the level of the user's preference,such as rating.Implicit feedback is made up of the record of user's behaviors,such as user's video viewing,product purchase history,popularity of items.Although it does not intuitively represent the user's preference,can reflect the focus of the user.Specifically,composed of user's behaviors,the temporal information can present the dynamics of the user's focus over time.The traditional recommender system utilizes explicit feedback to recommend a list of items,however,due to the sparsity and unavailability of explicit feedback,it results in biased representation of user profile.Since implicit feedback is comparatively rich in the volume,it is gradually earned the research interests from many scientists.However,the recommender system based on implicit feedback confronts three issues haste to be solved.Firstly,the scarcity of negative feedback makes implicit feedback challenge to be used.Secondly,because of the dynamics of user's preference,implicit feedback can change over time,so the static analysis of implicit feedback cannot fully excavate the regulation of data,which further reduce the prediction accuracy.Thirdly,due to the original denseness of implicit feedback,the optimized schema applied to the traditional recommender system does not suit implicit recommendation algorithm.And lacking of full consideration of whole data,such as observed data and unobserved data,can lower the accuracy of recommendation.This thesis focuses on the above three aspects as research objectives,and improves the prediction accuracy of personalized recommendation based on implicit feedback,temporal information,and matrix factorization.The contributions can be summarized as follows:(1)A recommendation algorithm TimeMF is proposed.The popularity of items can be treated as implicit feedback,and according to the analysis of the relationship between the popularity of items and time,we find popularity can dynamically change over time,and the stage of popularity can be split into two periods: popular period and unpopular period.Thus,we proposed a nonuniform weighting strategy based on dynamic popularity of items.The weighting schema solves the problem of the lacking negative feedback by assigning a non-uniform weight to missing value,which can dynamically extract the negative feedback from missing data.Meanwhile,to obtain the optimal solution,element adaptive gradient descent(AdaGRAD)is adopted to optimize objective function.AdaGRAD can assign an independent learning rate to each latent factor,and the learning rate can change over the gradient of latent factors.Through comprehensive experiments on Amazon movie dataset,the experimental results reveal that TimeMF,which adopts AdaGRAD and non-uniform weighting strategy based on dynamic popularity,can improve the accuracy of recommendation and outperform other state-of-the-art algorithms.(2)The recommendation algorithm TimeMFAW is proposed to make one step further on the basis of TimeMF.To sufficiently utilize observed data and simplify the optimized schema of TimeMF,we propose some solutions to deal with these issues,respectively.To overcome the shortcoming that the learning rate influences prediction accuracy,algorithm TimeMFAW adopts Element-wise Alternating Least Squares(EALS)which is an optimal schema to replace AdaGRAD.Meanwhile,we propose a non-uniform weighting strategy based on whole data,it can assign a non-uniform weight to observed data and missing data,respectively,which improves the coefficient of utilization of implicit feedback.Through comprehensive experiments on Yelp dataset,the experimental results reveal that the algorithm TimeMFAW can deal with the defects of algorithm TimeMF,and it outperforms other state-of-the-art algorithms.