Reproduce “Improving Recommendation Fairness via Graph Structure and Representation Augmentation” on OpenResearch

About this paper

Tongxin Xu, Wenqiang Liu, Chenzhong Bin, Cihan Xiao, Zhixin Zeng, Tianlong Gu

Fairness-aware Dual Data Augmentation (FairDDA) is introduced as a framework to enhance recommendation fairness in GCN-based systems by augmenting graph structure and feature representations. It achieves a 3.4% reduction in Demographic Parity and a 7.6% reduction in Equal Opportunity on MovieLens-1M, while preserving or slightly improving recommendation utility.

Key insights

• Addressing bias at the data level through targeted graph structure and representation augmentation offers a more fundamental and effective solution than solely debiasing the representation learning process.
• Principled hypotheses for identifying sensitive interactions and features enable precise data augmentation, mitigating bias without significantly distorting user preferences or disrupting graph integrity.
• Balancing the debiasing loss (HSIC) with graph reconstruction and contrastive learning objectives is crucial for simultaneously achieving high fairness and preserving recommendation utility.
• The proposed sensitive feature masking strategy is particularly effective on denser datasets, while sensitive edge pruning universally contributes to fairness without major utility compromise.

Summary by alphaXiv.