Research On Explainable Recommendation Algorithm Based On Graph Neural Networ

With the rapid development of online shopping and social media industries,the recommendation system has become one of the important tools of the Internet.Recommended purchases on shopping platforms,recommended songs on music platforms,and recommended friends on social platforms are all uses of recommendation systems.Accurately modeling user preferences based on their behaviors such as clicking,viewing,reading,and purchasing is at the heart of effective recommender systems.In the era of information explosion,the recommendation model aims to recommend items that users are interested in and solve the problem of information overload.Graph Neural Network(GNN)has been widely used in many prediction and recommendation tasks due to the feature structuring ability and collaborative ability of GNN for items and features.However,when the level of network increases,GNN produces over-smoothing problems that will degrade the recommendation performance.Most of the current recommendation models adopt deep learning paradigms,which lead to the urgent need for the explainability of recommendations.Explicable recommendations focus not only on producing accurate recommendation results,but also on producing persuasive explanations of how and why the item is recommended to a particular user.Most of the existing explainable review-based recommendation methods adopt a static independent method to extract the potential feature representation of user and item reviews,and express user preferences as static feature vectors,while users usually show different preferences when interacting with different items.Therefore,focusing on the key issues of the recommendation system,this paper proposes three more efficient recommendation models for the shortcomings of the current model,as follows:(1)First,in response to the problem of over-smoothing during model training.A collaborative filtering recommendation model based on GNN is proposed in this chapter to improve the recommendation performance.On the basis of using the graph neural network GNN,a collaborative filtering CF mechanism is added,which can solve The smoothing problem faced by existing recommendation algorithms using graph neural networks.This algorithm introduces the initial residual connection and identity mapping into the aggregation propagation process of the construction network to ensure that the network can perform deep learning and avoid the graph neural network from falling into over-smoothness after multiple convolutions,so as to improve the recommendation accuracy.(2)Next,for the potential feature representation of user and item comments,the extraction method based on the comment recommendation algorithm is more of a static independent method,with the user’s preference as the static feature vector,and in general,in different Among items,users show different preferences.Therefore,based on interactive attention,this chapter proposes an explainable recommendation method to analyze and study the correlation between user reviews and item reviews using interactive attention.Inspired by gate control in LSTM,the gating layer is added to the model to adaptively merge the feature vectors extracted from the two networks,and the attention factorization machine is used to further model the interaction of high-order features and achieve score prediction.Attention weight was used to measure comment information to improve the explainability of predicted scores explainability and the performance.(3)Finally,most recommendation systems adopt deep learning technology,which is relatively weak in terms of explainability.Moreover,most research on explainable recommendation systems is based on methods such as knowledge graphs or corpora from user queries and comments to generate explanations,which rely more on prior knowledge and are not conducive to the universality of the entire model.Establishing specialized network models for different items also consumes lots of time and money.Therefore,an explainable recommendation model based on graph neural network is designed,which introduces a multi-source heterogeneous graph attention network to model user behavior,and obtain recommended results based on the modeling results.The model does not require a lot of prior knowledge and is explained based on the attention value,so it has high explainability.In addition,in order to ensure the consistency of recommendations and explanations,a subgraph mining method is also introduced,which helps to bring a better experience to users.