Research Of Feature Fusion Sequential Recommendation

Sequential Recommendation predicts the user's next interactive item based on the user's historical interaction sequence.The user's historical interaction sequence contains rich feature information,which implies the user's interest and preference.Although some works have considered the contribution of these feature information and tried to use the feature information to improve the prediction effect,these attempts are all based on naive feature fusion methods,that is,the feature vector is used as auxiliary information and the item vector is spliced,and then input into model.Although these methods have achieved good results,they still have some drawbacks:(1)Missing implicit constraints of feature information.For the model designed for item sequence modeling,the feature information is used as auxiliary information for the item sequence to learn,and it is not possible to determine whether each feature has an impact on the item,or the influence of the feature information is implicit.The item sequence modeling model cannot capture the feature information well;(2)Vector space consistency and noise interference problems.The item vector and the feature vector are directly spliced into a new vector representation,which leads to the chaotic fusion of different vector spaces,introduces noise information,and destroys the consistency of the respective vector spaces of features and items.In order to solve these shortcomings,in this paper,we propose FAAN and MODAN model.The main contributions of this paper are as follows:1.Feature-Aware Attention Network FAAN Aiming at the problem that the existing sequence recommendation methods cannot effectively utilize the feature information,this paper proposes the FAAN model.We capture the transfer patterns implicit in each feature sequence by modeling each feature sequence independently.At the same time,a feature interaction module is proposed to explore the interaction between different features to capture user preferences more accurately.In order to enable the feature information to fully guide the model to capture user preferences,we innovatively fuse different feature information into the attention head,participate in the process of item sequence modeling,and provide explicit constraints for item sequence modeling.This paper conducts detailed experiments on the proposed FAAN model.Compared with the current mainstream sequence recommendation methods,the effects on the Movie Lens-1M and Ta-Feng datasets are(average)improved by 3.3%and 5.2%,respectively.2.Multi-Channel Orthogonal Decomposition Attention Network MODAN Based on the use of feature information,this paper proposes a MODAN model to solve the problems of vector space consistency and noise interference in existing sequence recommendation methods.The MODAN models item sequences and feature sequences differently in a multi-channel manner,thereby eliminating the noise interference between items and features and between features and features.We innovatively introduce orthogonal decomposition and reverse orthogonal decomposition to extract the contributions of items and features to each other to maintain the vector space consistency of item vectors and feature vectors,respectively.This paper conducts detailed experiments on the proposed MODAN model.Compared with the current mainstream sequence recommendation methods,the effects on the Movie Lens-1M,Amazon Apps and Diginetica datasets are(average)improved by 4.3%,8% and 6%,respectively.