Research On Heterogeneous Feature Fusion Recommendation Algorithm Based On Deep Learning

Many mobile applications use recommendation systems(RS)to help users make decisions such as where to visit,which products to buy,or which users to follow.The recommendations provided by these systems are undoubtedly of great significance to guide users to meet their needs.How to alleviate the poor recommendation effect caused by the sparsity of the traditional scoring matrix,consider adding comment text information,the deep learning method can be used to analyze the deeper preferences of hidden users and the hidden features of goods that have not been tagged can make up for the shortcomings of the scoring matrix.The research on heterogeneous feature fusion recommendation algorithm based on deep learning to provide personalized recommendation for users has become the starting point of this paper.First of all,aiming at the problem of how to extract features from the comment text,this paper proposes a feature extraction algorithm based on deep learning(XLNet-Caps).Considering that the review text is a valuable supplement to the user’s preferences and commodity attributes,the paper is usually unstructured and difficult to deal with.Therefore,the focus of XLNet＿Caps algorithm research is to mine user commodity features from unstructured review text.Then the extracted features are transformed into structured feature representations.Secondly,in view of the problem of how to improve the traditional Latent Factor Model(LFM)algorithm with poor recommendation effect,we choose to add scoring trust on its basis.User trust and item trust can be used to indicate that the user or product is rated.The degree of trust in the matrix,adding scoring trust to the traditional LFM algorithm can improve the performance of the recommendation.Thirdly,aiming at how to integrate the features from different data sources,this paper proposes a heterogeneous feature fusion recommendation algorithm based on deep learning(HTM),which combines the features extracted from the comment text of the score matrix respectively,including user and commodity features.Low-order features are extracted by FM and high-order features are extracted by CNN.Finally,low-order features are fused into the full connection layer for fractional prediction.Finally,based on the method proposed above,we do several groups of experiments on the real data set,using different evaluation indicators,comparing the parameters to find the best parameters,and comparing many classical or current advanced recommendation models,the algorithm proposed in this paper is effective.