Research And Optimization Of Attribute Information In Matrix Factorization Of Personalized Recommendation Algorithm

In big data era nowadays,the Personalized Recommendation technology can effectively improve user experience and solve the problem of information overload.Through analyzing in depth this big data,we can accurately portray the user models,recommend to him the users' subconscious items,and bring great conveniences to users and businesses.In addition,the attribute information as the real and effective information,which can accurately depict users and objects.Therefore,the attribute information can improve the recommendation accuracy.In the paper,the problems of initialization and cold start of matrix factorization in the recommended domain are studied by integrating attribute information.The problem of initialization in the matrix factorization:the traditional way is that the feature matrix is initialized using a certain range of random values,which will easily cause the problems that the prediction results maybe fall into local optimal solution and reducing the convergence speed in the iterative process.In order to overcome these problems,we propose two kinds of model to initialize the feature matrix of the property information:1)The initialization model based on attribute mapping,the attribute vectors are initialized directly using the item'attribute,the marked attributes of items are regarded as explicit features,and the rest of attributes are regarded as implicit features,and the user features are learned by the mapping mechanism;2)The initialization model based on the automatic coding of the deep neural network.In order to speed up the calculation efficiency and change the Invariance of feature dimension in the first way.In this paper,the automatic coding technique is used to obtain the abstract and low dimension features of the property,then;the SVD++is initialized with this feature.The cold start problem of Recommendation system:This paper mainly solves the problem of user cold start.For new users in the recommendation system,they have no historical score,so we can not recommend anything for them.In this paper,the attribute information is used as the additional information of users.First,the associations among users' attributes are obtained.We propose two models are as following:1)The matrix factorization model based on the attribute bias.According to the user's attribute bias information,the user's attribute bias is incorporated into the matrix decomposition,and the cold start user uses the attribute bias information and the global bias information to obtain the recommended results.2)The matrix factorization model based on the neighborhood's attributes.Firstly,the similarity of user attributes is calculated by means of semantic analysis,then,the decision tree is used to group the users to find the nearest neighbor in the group,finally,we construct the matrix decomposition model by using the nearest neighbor.The experiment shows that two models that use attribute information to initialize the feature matrix are superior to the traditional methods in both the recommendation accuracy and the convergence efficiency.The initialization method based on the attribute mapping is the highest,and the initialization method based on the automatic coding neural network is slightly lower than the former,however,the running time is decreased 50%.The problem of attribute sparsity is solved to a certain extent.Two models of using properties to solve cold start problem have been proposed,which alleviates the cold start problem of users to a certain extent and is better than the traditional cold start solution model.