Adaptive Thresholding for Multi-Label Classification via Global-Local Signal Fusion

TL;DR

This work tackles multi-label classification under extreme label imbalance by introducing a differentiable adaptive thresholding layer that fuses global rarity signals $IDF_l$ with local context signals $KNN_l(x)$ to produce per-label, per-instance penalties. The method defines a threshold $\theta_l(x)$ and optimizes a loss that penalizes logits near or beyond this threshold, using a combination of BCE with logits and a margin term. Empirically, on AmazonCat-13K, the adaptive thresholding approach achieves a macro-F1 of $0.1712$, outperforming baselines and prior state-of-the-art methods with a lightweight, modular architecture. The approach offers interpretability through explicit global and local components and demonstrates strong calibration and convergence properties, with potential extensions to full TF-IDF modeling and clinical NLP applications.

Abstract

Multi-label classification (MLC) requires predicting multiple labels per sample, often under heavy class imbalance and noisy conditions. Traditional approaches apply fixed thresholds or treat labels independently, overlooking context and global rarity. We introduce an adaptive thresholding mechanism that fuses global (IDF-based) and local (KNN-based) signals to produce per-label, per-instance thresholds. Instead of applying these as hard cutoffs, we treat them as differentiable penalties in the loss, providing smooth supervision and better calibration. Our architecture is lightweight, interpretable, and highly modular. On the AmazonCat-13K benchmark, it achieves a macro-F1 of 0.1712, substantially outperforming tree-based and pretrained transformer-based methods. We release full code for reproducibility and future extensions.