Research On Collaborative Filtering Recommendation Based On Deep Learning

With the development of data technology,personalized recommendation algorithm has gradually become the focus of research.Among personalized recommendation algorithms,collaborative filtering is used most.In recent years,in the development of machine learning,traditional collaborative filtering methods have been gradually combined with new technologies,some recent work has begun to apply deep learning methods,but mainly to model ancillary information.When modeling key information,the matrix factorization method is still used,and the implicit characteristics of users and items are still represented by inner products.Therefore,in the process of embedding,the synergistic signals implied in the interaction cannot be encoded effectively.In collaborative filtering,the embedding representation of users and items is a key issue.However,the past and present approaches have been to directly use existing features for embedding representation.In collaborative filtering,the embedding representation of users and items is a key issue.However,the past and present approaches have been to directly use existing features for embedding representation.In the process of embedded coding,the coding information hidden between the user and the project cannot be obtained,so the interactive information cannot be fully utilized,and finally the collaborative filtering recommendation quality is affected.To obtain a more accurate and comprehensive recommendation is the purpose of the recommendation algorithm,so it is very important to use the interactive information and auxiliary information of the user project for modeling.In the past research methods,it is regarded as a supervised learning problem.This problem separates the user,the project,and the interaction as entities,and then encodes the existing edge information.However,there are other higher-order connections between user,project,and entity relationships,and ignoring these connections makes the synergistic signals that can be obtained from the interaction behavior limited.That is,these methods ignore the higher-order connection between the user and the project's ancillary information,and therefore ignore important relevant information.The knowledge graph is used to decompose users,projects and their attribute relationships into independent entities,and connect the projects with one or more attribute relationships to obtain these higher-order relationships,thus improving the accuracy of recommendation.Aiming at the above problem of insufficient explicit expression,a new method of Attentive Neural Collaborative Filtering(ANCF)is proposed.The attention mechanism is used to distinguish the importance of neighbors,and each node is parameterized into a vector.We learn embedding representations by designing the attention dissemination layer.Use the attention mechanism to assign different weights,spread the embedding vector on the attention dissemination layer,learn and update the embedding vector of users and items,and explicitly express the coded information between users and items to realize embedded information explicit expression.On this basis,the knowledge graph is introduced,and a method Knowledge Graph Attention Collaborative Filtering(KGACF)is proposed.This method explicitly expresses the high-order connectivity in the knowledge graph in an end-to-end manner.By using the knowledge graph to obtain the high-order relationship of the auxiliary information of the user's items,it recursively propagates the embeddings from the node's neighbors(which can be users,items,or attributes).By superimposing multiple attention dissemination layers to explore high-order connectivity information,it can obtain more high-order relationships of auxiliary information.The ANCF and KGACF proposed in this paper are conceptually superior to the existing recommendation methods based on knowledge graphs or attention mechanisms.The empirical results on two datasets show that the KGACF method is better than the ANCF method,and the KGACF method is significantly better than the latest method.Further research has proved that the introduction of the attention mechanism is convenient for enhancing the embedding representation,and the introduction of knowledge graph embedding propagation is effective for high-order relationship modeling.