Research On Recommendation Algorithm Based On Graph Representation Learnin

Recommendation systems are an effective solution to the information overload problem,and many researchers learn user and item representations from user-item interaction data to predict user preferences.However,traditional recommendation methods often suffer from data sparsity and algorithm cold-start problems,leading to biased recommendation results.To address this problem,a large number of recommendation models often treat user and item attributes as collateral information to alleviate sparsity and improve recommendation performance.However,most of the models perform the recommendation task in a supervised learning mode,yet the observed user-item interactions are very sparse,making it insufficient to learn high-quality representations.Moreover,since most of the feedback provided by users is implicit rather than explicit,the observed interactions usually contain noise,which affects the performance of recommendation systems.Classical recommender systems are mainly concerned with accurately estimating user preferences,however,in recent years,the issue of fairness in recommender systems has attracted widespread attention.Unfairness may lead to negative consequences,which may have significant social impact.Therefore,it is crucial to consider fairness in recommendations and build reliable decision systems.Fairness machine learning aims to mitigate model bias in sensitive attributes,and among many existing fairness algorithms,counterfactual fairness measures the fairness of a model from a causal perspective by comparing the original data with the predictions of each individual in the counterfactual.Recently,some algorithms have extended counterfactual fairness to graph data,but most of them ignore the fact that sensitive attributes of each node’s neighbors may causally affect the prediction for this node,in addition to the fact that sensitive attributes may causally affect other features and graph structures.The graph data structure,due to its intuitive and flexible nature,can better represent the relationship between user items as well as collateral information.With its powerful ability to handle structured data and explore higher-order information,graph neural networks are widely used to learn node representations in graph data structures and have become a new advanced approach to many recommendation problems.Therefore,in order to solve the above mentioned problems in recommender systems,this paper introduces self-supervised learning and causal inference into knowledge-aware recommender systems to learn graph node representations,and uses the node representations learned by contrast learning in self-supervised learning and counterfactual learning in causal inference as auxiliary tasks to enhance existing knowledge-aware recommender systems,with the following main tasks:(1)knowledge-aware self-supervised representation learning recommendation algorithms.In this paper we propose a model for self-supervised contrastive learning on a hybrid structure of knowledge graph and user-item bipartite graph to solve the above problem.We design a knowledge-aware self-supervised graph contrastive learning model called KSGL,whose core idea is to efficiently learn user and item representations by pulling together augmented versions of the same user/item while pushing away augmented versions of different users/items.Specifically,KSGL first performs data augmentation on the input hybrid graph to generate multiple views of the target nodes,then refines the node embeddings in each view by graph convolutional networks(GCN),and finally updates the model by contrast loss.Through extensive experiments,we verify the effectiveness of KSGL,and the experimental results also show that KSGL can not only improve the accuracy of recommendations,but also obtain robustness to interactive noise.(2)Knowledge-aware counterfactual representation learning recommendation algorithm.In this paper,we introduce the concept of counterfactual fairness of graphs into knowledge-aware recommendation,which takes into account the sensitive properties of several points themselves and the bias caused by the sensitive properties of neighbors.To learn node representations for counterfactual fairness of graphs,we propose a knowledge-aware recommendation algorithm KCGR based on counterfactual data augmentation.in this algorithm,we generate counterfactuals corresponding to the perturbations of sensitive attributes of each node and its neighbors.Then,we enforce fairness by minimizing the difference between the representations learned from the original graph and the counterfactuals of each node.Finally,it is shown through experiments on real-world graphs that our framework outperforms state-of-the-art baselines in terms of counterfactual fairness of graphs,while also achieving comparable recommendation performance.