Research On Recommendation Algorithm Based On Deep Learning And Matrix Factorization

The development of network technology has promoted the generation of information platforms in various industries.At the same time,the widespread application of big data and cloud computing technologies have dramatically increased the number of users and items on the Internet.Under the above background,recommendation system has become an important tool to alleviate the information overload problem by helping users to obtain their interested items from massive information,which makes it an indispensable role for Internet applications.Matrix factorization is an important method for collaborative filtering recommendation,and widely favored by researchers of recommendation system,since it has good scalability and flexibility.In recent years,deep learning has achieved great success in many research fields,and its outstanding performance on feature representation has a great impact on information retrieval and recommendation systems.However,with the increasing number of users and items on the Internet,the sparseness,complexity,and uncertainty of data in the recommendation systems affect the accuracy of the recommendation results seriously.To deal with the aforementioned issue,this paper takes the advantages of deep learning and matrix factorization,and proposes three solutions.Specifically,the main research content and innovations of this article are as follows:1)Targeting the impact of data sparseness on the accuracy of recommendation,this paper com-bines deep learning technology with matrix factorization method and proposes a deep vari-ational matrix factorization recommendation algorithm.The algorithm designs and trainssemi-VAEs with a deep variational structure to capture user and item latent representationsrespectively.Further,it uses matrix factorization method with rating biased of user anditem to predict unobserved ratings.Finally,an optimization method is proposed for the jointmodel.This algorithm not only uses of the latent features extraction of the deep model,butalso retains the scalability of matrix factorization.It realizes the tight coupling of the deepmatrix factorization.2)Aiming at the problem of feature fusion in complex data and the high spatial complexitycaused by using deep learning in the factorization machine,this paper proposes a high-orderfactorization machine for recommendation,which designs a cross weights network to learnthe feature high-order interactions.Specifically,the algorithm designs a cross-weight net-work to learn the high-order interactions of features explicitly.Within the network,the crosslayers and compression layers are used to learn the weights of important high-order featurecombinations.The weight pooling layer is used to learn the feature interaction weights ofdifferent orders interactions to balance the impact of high-order and low-order interactionson the prediction results of the algorithm.The algorithm can better capture the inherentcorrelation of real data,since it considers not only the weights of different feature combina-tions but also the importance of different order interactions.At the same time,the explicitdeep cross operation avoids the unexplainable nature of deep neural network from learninhigh-order feature interactions.Its compression operation improves the space efficiency ofthe algorithm.3)To deal with the uncertainty of data in implicit recommendation,the paper models the miss-ing data from the perspective of causal relationship and proposes implicit recommendationalgorithm based on causal neural fuzzy inference.Fuzzy theory is used for the represen-tation the factors that affect the item exposure probability.Then,it combines with matrixfactorization to predict ratings.Considering that in actual applications,items observed byusers are often recommended by Internet platforms because of the application of the rec-ommendation systems.It affects the exposure of items.Based on the users and items,thisalgorithm analyzes two main factors affecting the exposure probability: the users' explicitpreference for the item or interact context,the popularity of the items.This paper usesfuzzy set theory to describe these two factors for their ambiguity,and uses neural fuzzy in-ference network to predict the exposure probability.In the neural fuzzy inference network,multi-layer perception is used to learn the weight of each fuzzy rule.Finally,it is combinedwith matrix factorization to predict user ratings for unobserved ratings and offer the usersinterested items.The experimental results show the rationality of the design and optimization of the deep variational matrix factorization algorithm,which have certain advantages in improving the accuracy of recommendation.At the same time,the importance of the cross-weight network is verified.It is proved that the high-order cross factorization machine algorithm is superior to other high-order factorization machine recommendation algorithms in accuracy.In addition,the comparison experiment results between the proposed algorithm causal fuzzy neural inference for implicit recommendation and other representative implicit recommendation algorithms prove the effectiveness of proposed algorithm.