Supervised Distributional Reduction via Optimal Transport and Dependence Maximization

The introduction of a novel algorithm, Supervised Distributional Reduction (SDR), aims to create target-aware representations by integrating optimal transport with explicit dependence maximization. SDR enhances the Fused Gromov-Wasserstein (FGW) objective, aligning the input distribution's relational structure with a collection of representative points. Additionally, it incorporates a direct dependence term that promotes the learned embeddings to effectively capture task-relevant signals. This approach tackles the issue of maintaining a balance between compression and predictive accuracy in supervised data reduction contexts. The research is available on arXiv, identified by the number 2605.27619.