Counterfactual reasoning can be effectively employed to perturb user-item interactions, to identify and explain unfairness in GNN-based recommender systems, thus paving the way for more equitable and transparent recommendations.

In this study, in collaboration with Francesco Fabbri, Gianni Fenu, Mirko Marras, and Giacomo Medda, and published in the ACM Transactions on Intelligent Systems and Technology, we explain and understand unfairness in recommendations generated by Graph Neural Network (GNN) models.

As recommender systems become increasingly sophisticated, their complexity often leads to a lack of transparency and potential biases. These biases can skew recommendations toward or against certain user groups, leading to unfair outcomes. Recognizing the critical need for fairness, we developed GNNUERS (GNN-based Unfairness Explainer in Recommender Systems), a framework that leverages counterfactual reasoning to identify user-item interactions contributing to unfair recommendations.

How GNNUERS works

GNNUERS operates by perturbing the user-item interaction graph in a memory-efficient manner. It aims to identify a set of interactions that, when altered, lead to a reduction in unfairness while maintaining the overall utility of the recommendations. The framework consists of several key components:

1. Bipartite Graph Perturbation. We adapt perturbation techniques to the bipartite nature of recommender systems, focusing on altering the interactions between users and items in a differentiable way. This approach is not only effective but also efficient, especially for sparse graphs.
2. Perturbed Graph Generation. Using a counterfactual model, GNNUERS predicts an altered relevance matrix by combining the normalized version of the perturbed adjacency matrix with the learned weights. This process is guided by a perturbation vector, updated during the learning process to ensure a monotonic trend in the number of perturbed edges.
3. Loss Function Optimization. The optimization of the perturbation vector is guided by a loss function that combines a fairness term based on Demographic Parity and a distance term controlling the perturbation extent. This function ensures that the framework minimizes unfairness while maintaining a close resemblance to the original graph structure.

Evaluation

We evaluated GNNUERS on three state-of-the-art GNN-based recommender models across four real-world datasets. Our experiments were designed to answer three critical research questions:

1. Can GNNUERS explain recommendation utility unfairness? Our findings demonstrate that GNNUERS can effectively narrow the distribution of unfairness, significantly reducing disparities in recommendation utility between different demographic groups.
2. Does GNNUERS minimally affect the recommendation utility of the protected group? The results confirm that while GNNUERS reduces utility for the unprotected group, it does so with a negligible loss for the protected group, thus maintaining the overall utility of the recommendations.
3. Can GNNUERS reveal the nature of unfairness through topological graph properties? By analyzing the user nodes connected to the perturbed edges, GNNUERS provides insights into how differences in graph properties between demographic groups contribute to unfairness.

Conclusion

Fairness in recommender systems is not just a technical challenge; it’s a fundamental aspect that affects user trust and satisfaction. With GNNUERS, our goal was understanding and mitigating unfairness in GNN-based recommender systems. Our framework provides a tool for identifying and explaining unfairness, paving the way for more equitable and transparent recommendations.

As our next goal, we want to extend the framework to handle more complex demographic subgroup structures and adapt it to a wider range of GNN architectures. Moreover, we aim to enhance the interpretability of the explanations provided by GNNUERS, making them more accessible and actionable for system designers and service providers.