Research On Feature Fusion Strategy Of Recommender Based On Deep Learning

With the rapid development of the Internet,the phenomenon of information overload is becoming more and more serious.How to quickly find the information users need becomes more and more difficult.Recommendation system can accurately recommend information to users,which is an important tool to solve this problem.Neural network is widely used in recommendation systems due to its powerful ability of feature extraction and feature modeling.However,when the features of neural network learning are fused with the traditional algorithms,the performance of the model will be limited if the features of users or items are fused with uniform weights.In order to solve this problem,this paper models the importance of the item features of autoencoder modeling and the high-order features of deep neural network modeling at the element level.It effectively solves the problem of insufficient fusion of neural network learning features and traditional algorithms.The main contributions of this paper are as follows:(1)Aiming at the autoencoder to model the features of the items,ignoring the problem that features have different importance at the element level,a new method HCF-DAE is proposed.Based on the hybrid recommendation algorithm of autoencoder and SVD++,this method models the importance of item features at the element level.This method introduces an attention mechanism to solve the problem of unified weight modeling of items features,uses SGD to learn weight parameters,screens out useful item feature,and improves the performance of recommendation.(2)For the high-level features of deep neural network modeling to predict the click-through rate of users to items,the contribution of each dimension of high-level features is different,a click-through rate prediction method based on implicit high-level feature importance modeling is proposed.This method uses the attention network to model the importance of each dimensions of high-order features,and model the weights of high-order features at the element level,which makes up for the lack of integration of the unified parameters of high-order features into the model.This method uses the attention mechanism to model the importance of high-order features,and flexibly integrates the attention network into different hybrid models to obtain two different recommendation methods to suit different recommendation situations.(3)In this paper,HCF-DAE and the click rate prediction method based on implicit high-order feature importance modeling are tested on several common datasets.After sufficient experiments,the proposed model is compared with the previous model,which proves that the proposed method can better integrate the features into the traditional recommendation algorithm,and improve the performance of recommendation algorithms.