Research On Multi-view Personalized Recommendation Method For Sparse Data

With the rapid development and fast changing of computer networks and the Internet business model,the Internet data has undergone an explosive growth.These data have the characteristics of large scale,high dimensionality and sparsity,which make it difficult and costly for users to extract valuable information.The problem of ‘information overload' is becoming increasingly serious than before.Fortunately,personalized recommendation is an important way to address this issue,which has been widely used in e-commerce and social networks etc.It recommends information and products that users might be interested in according to their personal information,interest and needs.Recommender systems fall into three categories according to the way of model construction.They are content-based method,collaborative filtering based method and hybrid recommendation method.Although great progress has been made in the study of various recommendation methods,there still exist a lot of limitations.Collaborative filtering recommendation is currently the most popular method among the above three ones,which has the advantages of simplified modeling and low data dependency,however,it still suffers from the problems of high data sparsity,poor expansibility,cold start and the difficulty of modeling user preferences.1)Data sparsity is one of the most serious problems of collaborative filtering recommendation.Collaborative filtering usually generates recommendation list relying primarily on the user-item interaction matrix(also known as the user-item rating matrix),and this matrix is usually extremely sparse.For example,the sparsity of Netflix dataset is about 99%,and the Movielens-10 M dataset contains approximately 98.7% unknown items.Therefore,many collaborative recommender systems might have poor recommendation quality if they merely depend on the interactive matrix to obtain recommendation results,since they cannot effectively filter out items satisfying users' needs.2)Influenced by the long tail effect,collaborative filtering also has the problem of modelling users with extreme difficulty and high complexity.Users' rating distribution is seriously unbalanced,that is,a small number of users might have more ratings on items,and while most users might have only a few ratings.This makes it difficult for recommender systems to grasp user preferences and user interest accurately,which would lead to over-fitting or under-fitting problems in user preference modeling.3)In addition,there are some other problems such as poor scalability and cold start in collaborative recommender systems.With the development of machine learning,especially deep learning,the utilization patterns of various complex features have gradually matured.The construction of recommender systems through analysis,modelling and feature extraction of multi-source data based on machine learning technologies,is the current research hotspots.In this paper,we focus on multi-view personalized recommendation model for sparse data in view of the above issues.Firstly,we study the low rank matrix completion algorithms to fill in missing values in user-item interaction matrix,to solve the problem of data sparsity.Secondly,we combine user-item interaction matrix with multi-view information under deep learning framework when modelling user preferences,to solve the difficulty in user preference modelling and the cold start problem.The main research contents and contributions of this paper are as follows:1)A collaborative filtering method integrating robust non-negative matrix factorization/completion and subgroups partitioning.Many traditional matrix factorization methods have poor interpretability because of the negative values in their factorization results.In the real world,many applications require the data to meet non-negative constraints,and in addition,the user-item rating matrix is usually accompanied by noise and outliers.Therefore,we put forward a collaborative filtering algorithm combining low rank and robust nonnegative matrix factorization/completion with subgroups partitioning(abbreviated as LR-RNMFC),to solve the above problems.In the low rank matrix completion stage,we build low rank matrix filling model based on robust non-negative matrix factorization,and deduce an efficient iterative method to solve the model and provide the proof of convergence.The algorithm not only uses the reconstruction matrix to fill in the missing items,but also obtain the robust non-negative decomposition expression of the scoring matrix.In collaborative filtering recommendation stage,by exploiting the clustering property of the non-negative submatrix,we divide the reconstruction solution of the original matrix into user-interest subsets based on block model clustering method,resulting in collaborative recommendations for target users.Empirical results reveal that the proposed method can fill the missing items of the target matrix effectively,and the recommendation accuracy of traditional algorithms can be improved significantly when combined with our LR-RNMFC.It indicates sufficiently that the matrix completion is an effective way to solve the sparsity problem of rating matrix.2)A stochastic sub-gradient method for low-rank matrix completion of collaborative recommendation.Being the main data source for the construction of recommender systems,the user-item interaction matrix has the characteritcs of high dimension,data sparsity and extremely uneven distribution,which brings challenges for learning user preferences.We focus on nuclear norm regularized matrix completion model in large matrices,and propose a new model named stochastic sub-gradient method for low rank matrix completion(SS-LRMC).To the problem of traditional SVT algorithm that would use one fixed threshold to shrink all singular values during iterations,and the enormous computation burden when faced with large matrices,we define an adaptive singular value thresholding operator,and put forward a kind of matrix completion model applicable for user-item rating matrix of collaborative filtering.During iterations,we combine stochastic sub-gradient descent techniques with the adaptive singular value thresholding operator to obtain low rank intermediate solutions.Empirical results confirm that our proposed model and algorithms outperform several state-of-the-art matrix completion algorithms and the application to collaborative filtering recommendation can effectively alleviate the sparseness problem of the user-item rating matrix and can significantly improve recommendation accuracy.Moreover,complexity analysis shows that it is easy for us to extend the proposed algorithm to application scenarios with larger data sets.3)A wide and deep model based on multi-source information-aware recommendation is proposed.Although matrix completion algorithms can effectively alleviate the sparsity of rating matrix,they customarily model the linear feature interactions between users and items,instead of the complex nonlinear structures.In order to better depict user preferences and item features,we deepen the linear model of LR-RNMFC,to establish a wide and deep model,which we named WDMMA,based on multi-source information of user-item interaction matrix,attribute and context.The wide part mainly handles the linear interactions between user and items,while the deep part portrays the high-order nonlinear interactions.We pre-train the wide as well as the deep part of WDMMA using LR-RNMFC in the embedding layer.Upon the embedding layer is the pooling layer,we define a pooling operation named AC-pooling,which is used to model various low-order interactions among users,items,attributes and context information.Upon the pooling layer we stack some hidden layers to capture the high-order nonlinear feature interactions.Experiments on two public datasets show that,it is an effective way to consider both linear user-item interactions and higher-order nonlinear interactions fusing multi-source information.This method can learn the complex nonlinear feature patterns successfully and effectively,and is helpful to improve recommendation performance.4)The method of attention-based context-aware sequential recommendation using GRU.In addition to context and attribute information,sequence information plays an increasingly important role in recommendations.To the problem that sequential recommendation methods give undue importance to sequence changes and insufficient emphasis on the correlation between adjacency items,we consider that the recommendation list cannot be totally affected by users' recent behaviors.We emphasize on the continuity of user interest when modelling user preference,and propose the model ACA-GRU.In ACA-GRU,we divide the context into four types,including input context,correlation context,static interest context and transition context.By redefining the update gate and reset gate of GRU unit,we calculate the global sequential state transition of RNN determined by these four types of context information,to model the dynamics of user interest.In order to solve the problem of outliers in sequence recommendation,we utilize attention mechanism to distinguish the importance of each item in the rating sequence by calculating the correlation context.The impact of outliers that are less informative and predictive is decreased in this process.Experimental results show that the performance of ACA-GRU is superior to state-of-the-art context-aware as well as sequence recommendation algorithms,which indicates the effectiveness of attention mechanism in sequence recommendation.