| In recent years,the amount of data generated globally every year has shown explosive growth.In order to efficiently and accurately obtain valuable information from massive data,recommender systems emerge as the times require.In the operation of the recommendation system,the knowledge map has shown its strong potential as a means of supplementing data.The recommendation system can obtain rich entity and relationship information through the knowledge graph,which can help it better model the relationship between users and items,which not only helps to reveal various correlations between items,but also Can be used to explain user preferences.Therefore,introducing a knowledge graph into a recommendation system can not only improve the accuracy of the recommendation,but also improve its interpretability.With the emergence of the graph neural network,due to its advantages in learning graphstructured data,many studies have introduced the graph neural network into the recommendation system to better obtain information in the knowledge graph.A recent technological trend is to design end-to-end models based on information dissemination methods,however,existing partially dissemination-based methods fail to capture the highorder collaboration signals of projects while ignoring the information hierarchy in knowledge graphs.In order to deal with these problems,this paper proposes a recommendation method based on dual embedding aggregation,here referred to as DEIA,which models user-item representation and knowledge items in hyperbolic space,expresses collaboration signals and knowledge associations respectively through dual information embedding,and utilizes information Gated aggregation to discriminate two kinds of information to better capture user behavior patterns.After multiple experiments on the three data sets of Book-Crossing,Movie Lens-1M and Last.FM,this paper shows that this method has better results than other comparison methods.But there is still the problem of over-smoothing interacting with noise caused by node degree skew.In order to alleviate the existing problems in DEIA,this paper proposes a differentiable sampling method DSKG based on knowledge graph.It learns the relevant distribution of connected items from the knowledge graph,and samples suitable items for recommendation according to this distribution.In this paper,a differentiable sampling strategy is designed to enable the selection of relevant items to be jointly optimized with the model training process.By replacing the uniform sampling method used by DEIA with the differentiable sampling method DSKG,the algorithm model DEIA+DSKG in this paper is obtained,which is a differentiable knowledge map recommendation model based on dual embedding aggregation.The experimental results in three data sets show that the improvement effect of DEIA+DSKG is very significant.The experimental results show that the differentiable knowledge graph recommendation model based on dual embedding aggregation(DEIA+DSKG)proposed in this paper can better capture the high-order synergy signals and the hierarchical structure of information,and at the same time,it can effectively alleviate the excessive smoothing and Noise interaction problem,showing the effectiveness of the recommendation method. |
