Recommendation Model Based On Knowledge Graph And Graph Neural Network

Recommendation system is an important means for users to obtain effective information in various domains.Traditional recommendation algorithms(e.g.,collaborative filtering)rely on historical interaction data between users and items.However,the lack of sufficient interaction data leads to severe data sparsity and cold start problems.This thesis introduces Knowledge Graph(KG)as auxiliary information for recommendations to solve these problems.KG not only reveals the correlation between items(e.g.,two movies are starred by the same actor),but also can be used to explain user preferences(e.g.,a user chooses to watch a movie because of a certain actor).However,existing knowledge graph-based recommendation methods cannot represent user and item features at a fine-grained level,suffering from the sparse supervised signal problem,which may lead to a drop in recommendation accuracy.In addition,most recommendation methods only focus on one type of user’s behavior(e.g.,purchase)and overlook the diversity of user behavior.This thesis mainly uses graph neural networks to overcome these limitations of existing knowledge graph-based recommendation methods.The primary research and innovation points of this thesis are discussed in the following parts:(1)A recommendation model fusing knowledge graph and user intents is proposed.Firstly,the model incorporates the attention mechanism to calculate the impact of different attribute information on user preferences.Moreover,a heterogeneous graph neural network is employed to learn the features of items.Considering the user’s intent to select an item,the relationship between user and item attributes is computed through the attention mechanism to learn the user’s preferences at a finer granularity.Meanwhile,each item entity is embedded into different relational attribute spaces under different layers of the GNN framework,thereby effectively avoiding semantic pollution caused during aggregation.Finally,the recommendation result is obtained through the inner product of item representation and user representation.By conducting experiments on public datasets in movies,music and books,the results demonstrate that the model proposed in this thesis improves on all metrics compared toother baseline models,and improves the accuracy and interpretability of recommendations.(2)A recommendation model based on contrastive learning is proposed.To address the problem of supervised signal sparsity caused by extremely lack of user-item interactions,this thesis adds a contrastive learning module based on the first part.The module uses the idea of contrastive learning to divide the original data into two views: a user-item interaction view and an item-attribute semantic view.The item representations are learned by Light-GCN in the interactive view,which performs contrastive learning between them and the item representations learned by the model of the first part.This approach mines the differences between different item representations in a self-supervised manner,effectively alleviating the supervised signal sparsity problem.Experimental results show that the addition of the contrastive learning module can alleviate the supervised signal sparsity problem and further improve the recommendation performance.Compared with the optimal model,the maximum increase of Recall@15 and Precision@15 in each data set is 1.84% and 2.42%.(3)A multi-behavior recommendation model based on user relationships is proposed.The model assumes that users with similar preferences are likely to have similar interactions.To capture user features,user-user relationships are added to the original user-item and item-item data.The model utilizes item-item data,user-user,and user-item data from different behavior types as inputs to the network,which learns the characteristics of users and items with graph attention networks.To fuse the features under different behaviors,the gated mechanism is utilized.Finally,the obtained user features and item features are fused to predict the probability of interaction between the user and the item for the target behavior.Through experiments on two datasets,the model outperforms other baseline models in the HR@10 and NDCG@10evaluation metrics.