Exploiting Conjugate Label Information for Multi-Instance Partial-Label Learning

TL;DR

This work tackles dual inexact supervision in multi-instance partial-label learning by introducing EliMipl, which leverages conjugate label information from both candidate and non-candidate label sets. A scaled additive attention mechanism produces bag-level representations, while a tri-component loss (mapping $L_{ma}$, sparsity $L_{sp}$, and inhibition $L_{in}$) encodes the label-space structure and enforces sparsity toward the true label. Empirical results on diverse benchmark and real-world datasets show that CLI-driven EliMipl outperforms existing MIPL methods and PLL baselines, particularly as the number of false-positive labels grows. The proposed approach delivers both higher accuracy and interpretable attention patterns, highlighting its practical impact for weakly supervised learning in complex label spaces.

Abstract

Multi-instance partial-label learning (MIPL) addresses scenarios where each training sample is represented as a multi-instance bag associated with a candidate label set containing one true label and several false positives. Existing MIPL algorithms have primarily focused on mapping multi-instance bags to candidate label sets for disambiguation, disregarding the intrinsic properties of the label space and the supervised information provided by non-candidate label sets. In this paper, we propose an algorithm named ELIMIPL, i.e., Exploiting conjugate Label Information for Multi-Instance Partial-Label learning, which exploits the conjugate label information to improve the disambiguation performance. To achieve this, we extract the label information embedded in both candidate and non-candidate label sets, incorporating the intrinsic properties of the label space. Experimental results obtained from benchmark and real-world datasets demonstrate the superiority of the proposed ELIMIPL over existing MIPL algorithms and other well-established partial-label learning algorithms.

Figures (5)

• Figure 1: (a) A multi-instance bag is labeled with a candidate label set $\mathcal{S}=\{k_1, k_2, k_5, k_7\}$. (b) The decomposition of the complete label matrix, where $m$ and $k$ represent the number of multi-instance bags and categories, respectively.
• Figure 2: Predicted probabilities of DeMipl (left) and EliMipl (right) on the sample in CRC-MIPL-Row dataset.
• Figure 3: The pipeline of EliMipl, where $\mathcal{L}_{\text{ma}}$, $\mathcal{L}_{\text{sp}}$, and $\mathcal{L}_{\text{in}}$ refer to mapping loss, sparsity loss, and inhibition loss, respectively.
• Figure 4: The classification accuracies of EliMipl, DeMipl, and MiplGp on the Birdsong-MIPL dataset with varying $r$.
• Figure 5: Attention scores for a test bag. Red and blue are the attention scores of positive and negative instances, respectively.