DPAA outperforms SOTA at debiasing GNN-based recommenders
Adaptive weighting in message passing surfaces long-tail items without sacrificing accuracy.
Graph neural network (GNN) collaborative filtering models are powerful but suffer from popularity bias: repeated message passing amplifies popular items while suppressing long-tail ones. Existing debiasing approaches—such as re-weighting, regularization, causal methods, and post-processing—fall short in GNN-based settings because they don't directly counteract bias propagated through the aggregation process. Recent in-aggregation weighting methods rely on static heuristics or unstable embedding estimates, limiting their effectiveness.
To address this, the authors introduce DPAA, which integrates adaptive, embedding-aware interaction weighting and layer-wise weighting directly into message passing. Interaction-level weights are derived from a representation-aware popularity signal, stabilized by a smooth transition from pre-trained to evolving model embeddings during training. Layer-wise weighting amplifies higher-order neighborhoods to surface diverse and underexposed items. Experiments on real-world and semi-synthetic datasets show DPAA outperforms state-of-the-art popularity-bias correction methods for GNN-based collaborative filtering.
- Popularity bias is amplified by repeated message passing in GNNs, suppressing long-tail items.
- Existing debiasing methods (re-weighting, causal, post-processing) fail in GNNs as they don't counter bias in aggregation.
- DPAA uses adaptive embedding-aware weights and layer-wise weighting to promote diverse items, outperforming SOTA on benchmarks.
Why It Matters
Helps recommender systems surface niche content, improving user experience and fairness for lesser-known items.