LTGA: Learning Per-Edge Tsallis Index for Graph Attention
A new research paper on arXiv (2608.02938) introduces LTGA (Learnable Tsallis Graph Attention), a graph attention layer that learns the Tsallis entropic index q jointly with network weights. This allows attention to interpolate between heavy-tailed (q<1), softmax (q=1), and compact-support (q>1) distributions, addressing the need for different attention shapes in homophilic and heterophilic graphs. The method operates at four granularities, from a global scalar to a per-edge index, with a bounded reparameterization that initializes models at the GAT baseline. Experiments across eight benchmarks with ten seeds show that LTGA-Edge achieves the best average rank (2.75), though the omnibus test does not reject the null hypothesis (p=0.199). Notably, learning q does not outperform a validation-tuned frozen grid, which reaches 61.4% accuracy, while tuned alpha-entmax achieves 62%.
Key facts
- LTGA learns the Tsallis entropic index q jointly with weights.
- Attention shapes range from heavy-tailed (q<1) to compact-support (q>1).
- Four granularities: global scalar to per-edge index.
- Bounded reparameterization starts at GAT baseline.
- LTGA-Edge achieves best average rank 2.75 across eight benchmarks.
- Omnibus test p-value 0.199, not significant.
- Validation-tuned frozen grid reaches 61.4% accuracy.
- Tuned alpha-entmax achieves 62% accuracy.
Entities
Institutions
- arXiv