[Paper Review] Transferable Fairness for Cold-Start Recommendation

Yunqi Li, Dingxian Wang+2 more|arXiv (Cornell University)|Jan 25, 2023

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes a transfer learning framework, TFR, to transfer counterfactual fairness and recommendation performance from warm users (source domain) to cold-start users (target domain). By integrating adversarial learning to disentangle sensitive attributes from user embeddings, TFR achieves high fairness and improved recommendation accuracy on cold-start users without requiring target user interaction data.

ABSTRACT

With the increasing use and impact of recommender systems in our daily lives, how to achieve fairness in recommendation has become an important problem. Previous works on fairness-aware recommendation mainly focus on a predefined set of (usually warm-start) users. However, recommender systems often face more challenging fairness issues for new users or cold-start users due to their insufficient amount of interactions. Therefore, it is essential to study whether the trained model still performs fairly for a new set of cold-start users. This paper considers the scenario where the recommender system meets new users who only have limited or even no interaction with the platform, and aims at providing high-quality and fair recommendations to such users effectively. The sufficient interaction data from warm users is treated as the source user domain, while the data from new users is treated as the target user domain, and we consider to transfer the counterfactual fairness from the source users to the target users. To this end, we introduce a framework to achieve transferable counterfactual fairness in recommendation. The proposed method is able to transfer the knowledge of a fair model learned from the source users to the target users with the hope of improving the recommendation performance and keeping the fairness property on the target users. Experiments on two real-world datasets with representative recommendation algorithms show that our method not only promotes fairness for the target users, but also outperforms comparative models in terms of recommendation performance.

Motivation & Objective

• Address the gap in fairness-aware recommendation for cold-start users, who lack sufficient interaction data.
• Transfer fairness properties—specifically counterfactual fairness—from warm users (source domain) to cold-start users (target domain).
• Maintain high recommendation performance for cold-start users while ensuring fairness through knowledge transfer.
• Develop a method that does not require interaction data from target users, enabling unsupervised transfer of fairness and utility.

Proposed method

• Use a source domain of warm users with abundant interaction data to train a counterfactually fair recommender model.
• Integrate adversarial learning to make user representations invariant to sensitive attributes (e.g., gender), ensuring counterfactual fairness.
• Train the recommendation model and adversary jointly to disentangle sensitive features from user embeddings while preserving recommendation utility.
• Transfer the fair model from source to target users via domain adaptation, without fine-tuning on target user data in the unsupervised setting.
• In the supervised setting, incorporate limited target user interactions to further improve recommendation performance while preserving fairness.
• Use an adversary to evaluate fairness by measuring its ability to predict sensitive attributes from user embeddings; low AUC indicates high fairness.

Experimental results

Research questions

• RQ1 Can counterfactual fairness learned from warm users be effectively transferred to cold-start users without access to their interaction data?
• RQ2 Does transferring fairness knowledge from source to target users improve recommendation performance for cold-start users?
• RQ3 How does the proposed method compare to baselines in maintaining fairness and recommendation accuracy under both unsupervised and supervised settings?
• RQ4 To what extent does adversarial disentanglement of sensitive attributes from user embeddings ensure counterfactual fairness in cold-start scenarios?

Key findings

• TFR achieves AUC scores around 0.5 on adversarial attacks for sensitive attribute prediction, indicating that user embeddings are nearly independent of sensitive features, thus ensuring counterfactual fairness.
• In the unsupervised setting, TFR outperforms baselines trained only on source data and most baselines trained on target data alone, demonstrating effective transfer of recommendation performance.
• In the supervised setting, TFR surpasses all baseline models across all evaluation metrics, even those trained on combined source and target data, confirming its superiority in performance and fairness.
• The model maintains high recommendation quality (measured by AUC and other metrics) while achieving strong fairness, as evidenced by low AUC in adversary classification.
• TFR performs comparably to full-data baselines in recommendation performance despite not using target user interaction data, proving the effectiveness of unsupervised transfer.
• The supervised variant of TFR achieves better fairness and performance than the unsupervised variant, showing the benefit of incorporating limited target data when available.