Fairness in Federated Recommender Systems: A Comprehensive Survey

Federated recommendation systems have gained popularity in recent years due to their ability to personalize content without compromising user privacy. However, these systems often neglect fairness, which can lead to biased recommendations. This article provides a comprehensive overview of fairness-aware federated recommendation systems, including their challenges and solutions.

Challenges

1. Unfairness in conventional federated recommendation systems: Many prior works (Islam et al., 2019; Yao and Huang, 2017; Li et al., 2021a) mitigate unfairness in conventional recommendation systems, which call for exchanging private attributes with the server and compromising privacy in federated settings.
2. Local training for fairness: Contrary to this, training locally to achieve fairness in federated recommendation systems without exposing user demographic information becomes exceptionally challenging.
3. Graph structure in recommendation data: The backbone for existing federated recommendation methods is matrix factorization (MF), which only considers explicit user-item interactions for updating embeddings; the implicit intricate interaction in the form of a bipartite graph does not play any role.

Solutions

1. Graph attention networks (GATs): GATs pave the way to build a GNN-based recommendation system (Wu et al., 2022b). Due to their efficiency and inductive learning capability, GATs can utilize the graph structure in recommendation data to improve fairness.
2. Fairness-aware federated learning: Owing to its efficiency and inductive learning capability, GNNs can further utilize this graph structure in RSs data to achieve fairness without compromising privacy.
3. Inductive representation learning on large graphs: Hhardt et al. (2016) proposed an equality-aware algorithm for matrix factorization that considers the intricate structure of a bipartite graph, enabling inductive representation learning on large graphs.

Conclusion

Fairness-aware federated recommendation systems are essential to ensure privacy and fairness in personalized content recommendation. By utilizing graph attention networks and inductive representation learning on large graphs, these systems can overcome the challenges of conventional federated recommendation systems and provide more accurate and fair recommendations to users.