Decision Boundary-aware Generation for Long-tailed Learning

A new framework, Decision Boundary-aware Generation (DBG), addresses long-tailed data bias in machine learning. Long-tailed datasets skew decision boundaries toward head classes, reducing tail class accuracy. Diffusion-based generative augmentation and head-to-tail transfer have been used to mitigate this, but the latter can cause latent non-local feature mixing, leading to decision boundary overlap and tail class distribution shift. DBG identifies this boundary ambiguity problem and generates informative near-boundary samples to promote separable decision spaces. The framework rebalances long-tailed datasets and improves classifier performance. Results on standard long-tailed benchmarks show consistent improvements.