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

TL;DR

This paper proposes an adaptive thresholding method for multi-label classification that combines global and local signals, improving calibration and performance, especially under class imbalance and noise.

Contribution

It introduces a differentiable, fusion-based adaptive thresholding mechanism that enhances multi-label classification accuracy and calibration.

Findings

Achieves macro-F1 of 0.1712 on AmazonCat-13K

Outperforms tree-based and transformer-based methods

Provides a lightweight, interpretable architecture

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.