Attention-Based Causal Graph Convolutional Collaborative Filtering

Youhan Qi, Xinglin Liu, Chenyu Li, Ying Wang

Abstract: Graph-structured data can naturally represent complex relationships between entities, hence graph neural networks have been widely employed in collaborative filtering methods to facilitate data mining on user-item graphs. However, the use of traditional methods may be affected by confounding factors in the prediction process, resulting in unfairness or suboptimality in the recommendation, and reducing the performance of the recommendation system. Causal Inference, as a powerful method, facilitates understand the dependency relationships between variables and identify potential data biases, thereby providing more accurate and fair recommendations for recommendation systems. Therefore, we propose a fair collaborative filtering model called Attention-based Causal Graph Convolutional Collaborative Filtering (ACGCF), which can address bias issues in recommendation systems. Firstly, ACGCF preprocess the dataset of the recommendation system and then the graph convolution method and attention mechanism are used to model the higher-order relationship on the user-item graph. Secondly, the backdoor adjustment in causal learning is used to eliminate the influence of confounding factors, and the do-operations are adjusted according to the user’s feedback and behavior. Finally, integrate the user-user graph, item-item graph, and the causal debiasing user-item graph, and utilize graph matching to derive the final recommendation results. Furthermore, we have numerically verified the superiority of our method in terms of recommendation performance with other comparison methods.