Research On Personalized Recommendation Algorithm Based On Matrix Factorization

With the rapid development of the Internet,the amount of information on the Internet has increased rapidly,which brings convenience to users but also faces the problem of information overload,therefore,recommender systems have developed rapidly as an effective solution.In recent years,recommendation algorithms as the core of recommender systems have been studied extensively.Matrix factorization is a classic and important technique which is commonly used in recommendation algorithms,and it has received extensive attention and research so far.Traditional matrix factorization predicts ratings with the inner product on the latent features of users and items,but it often confronts the data sparsity problem.In recent years,deep learning has proven able to learn good feature representation in natural language processing,image classification,and so on,therefore a large number of researchers have introduced deep learning into recommender systems to solve the data sparsity problem.However,there are still the following problems.Firstly,existing methods do not fully combine rating matrix and side information,and cannot learn the joint feature representation of multiple data sources.Secondly,features extracted by autoencoder are different from those from rating prediction,which difference is not considered by existing recommendation algorithms to decrease recommendation performance.Finally,user preferences cannot be accurately represented by a single feature vector.Some depth research has been conducted in this thesis to solve the above problems faced by existing methods.The main research work of this thesis is as follows:1.This thesis deeply studies existing personalized recommendation algorithms and the personalized recommendation algorithms combined with deep learning,and conducts corresponding analysis and summary,to find out the shortcomings in existing research.2.This thesis proposes a new feature extraction model——Semi-SDAE based on autoencoder.Most of the existing feature extraction models can only deal with a single data source problem,and it is difficult to integrate multiple data source information.This thesis proposes a feature extraction model based on autoencoder to obtain richer feature representation.This model not only integrates multiple data source information,but also removes the limitation that the input data and output data of autoencoder must be equal,making the model more practical.3.This thesis proposes a self-adaptive matrix factorization recommendation algorithm——SAMF based on matrix factorization and combining with Semi-SDAE.Existing recommendation algorithms directly take the feature from autoencoder into rating prediction,but do not consider the difference of tasks of feature extraction and rating prediction.This thesis conducts related research and proposes a self-adaptive matrix factorization recommendation algorithm,making the extracted feature can be efficiently used to rating prediction.4.This thesis proposes a layered multi-granularity matrix factorization recommendation algorithm——LMGMF to solve the problem of low precision with a single latent vector based on matrix factorization.This algorithm not only overcomes the problem that single latent vector can not accurately represents users' preferences and items' features,but also overcomes the problem that existing recommendation algorithms based on deep learning predict ratings with only the last layer,but ignores the information loss caused by feature transformation in each layer of the neural network.5.Experimental results on real-world datasets show that recommendation algorithms proposed by this thesis perform better,and verify the effectiveness of the proposed algorithms.