Research On Recommended Method Based On Deep Learning And Feature Fusion

As an effective tool to solve the information overload of the modern Internet,the recommender system can provide users with personalized and accurate information sources and a decision-making basis,which is blending into the daily life of more and more people.Data sparsity is one of the main problems faced by the current recommender system.With the rapid development of in-depth learning,efficient feature extraction and effective fusion is a worthwhile research direction,which provides strong support for alleviating data sparsity.The purpose of this paper is to study how to integrate the non-linear modeling ability of deep learning with the traditional recommendation algorithm’s linear modeling ability,so that the algorithm combines the fast searching ability of the linear model with the generalization ability of the non-linear model,to alleviate the data sparsity problem of the recommender system.This paper will study the feature fusion technology of the model,and use different feature fusion methods in three places（model hidden layer,module output layer,and feature fusion module）to improve the recommendation performance of the model.The main work of this paper is as follows:1)Given the different characteristics of sparse category and dense continuous features,and considering the different roles they play in model operation,a feature extraction strategy based on structure separation is proposed.The two features are respectively input into the neural network to extract the features,and then the output features are fused.2)A hybrid recommendation algorithm,HR-MF-DL,based on deep learning and matrix factorization,is proposed to solve the problem that the matrix factorization algorithm produces fit when the score data is sparse.The algorithm uses a deep neural network to extract the implicit features of users and items,fuses them with a matrix factorization algorithm,and fuses them with point products to improve the recommended performance on real datasets.3)For the gradient diffusion phenomenon that occurs during deep learning,RRMLP is integrated with residual block,activation function,and normalization method.The comparison with multi-layer perceptron（MLP）in the experiment shows that RRMLP can better extract features and improve the recommendation performance.After applying RRMLP to the HR-MF-DL algorithm,the recommendation effect of the algorithm is further improved and verified on two real datasets.