Researchers have developed FairPGT, a novel partitioned graph transformer designed to address demographic bias in graph clustering. This model restricts attention mechanisms to specific partitions and uses routing tokens to maintain scalability, dividing the computational cost by the number of partitions. FairPGT incorporates a graph-aware self-training target and enforces fairness at two levels: a mutual-information term for demographic parity in cluster assignments and a normalized cross-group attention ratio (nCGAR) to guide attention toward group neutrality. Experiments across seven benchmarks demonstrate that FairPGT achieves superior clustering performance and statistical parity, significantly reducing bias compared to existing methods. AI
IMPACT This research introduces a novel approach to mitigate bias in graph clustering, potentially leading to fairer AI systems in applications involving sensitive data.
RANK_REASON The item is an academic paper detailing a new method for graph clustering. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- CatalyzeX
- CRediT
- DagsHub
- Digital Equipment Corporation
- FairPGT
- Gotit.pub
- Hugging Face
- IArxiv
- METIS
- Mon1
- Northern Mariana Islands
- Pokec-n
- Pokec-z
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →