Multi-Instance Partial-Label Learning with Margin Adjustment

TL;DR

The paper tackles margin violations in multi-instance partial-label learning by proposing MIPLMA, which integrates a margin-aware attention mechanism for instance-space margin adjustment and a margin distribution loss for label-space margins. By dynamically tuning margins via a temperature schedule and jointly optimizing $\mathcal{L} = \mathcal{L}_{\text{d}} + \lambda \mathcal{L}_{\text{m}}$, the approach improves disambiguation of true labels and separation between candidate and non-candidate labels. Empirical results on benchmark and real-world CRC-MIPL datasets show consistent, substantial gains over MIPL, PLL, and MIL baselines, including strong performance with deep features and across data-degradation settings. The work also demonstrates that margin adjustments extend benefits to MIL and PLL, highlighting the practical impact for weak supervision scenarios in diverse domains.

Abstract

Multi-instance partial-label learning (MIPL) is an emerging learning framework where each training sample is represented as a multi-instance bag associated with a candidate label set. Existing MIPL algorithms often overlook the margins for attention scores and predicted probabilities, leading to suboptimal generalization performance. A critical issue with these algorithms is that the highest prediction probability of the classifier may appear on a non-candidate label. In this paper, we propose an algorithm named MIPLMA, i.e., Multi-Instance Partial-Label learning with Margin Adjustment, which adjusts the margins for attention scores and predicted probabilities. We introduce a margin-aware attention mechanism to dynamically adjust the margins for attention scores and propose a margin distribution loss to constrain the margins between the predicted probabilities on candidate and non-candidate label sets. Experimental results demonstrate the superior performance of MIPLMA over existing MIPL algorithms, as well as other well-established multi-instance learning algorithms and partial-label learning algorithms.

Figures (7)

• Figure 1: Margin violations in the instance space and the label space. (a) and (b) depict the attention scores of EliMipl and MiplMa for the same test bag in the FMNIST-MIPL dataset. Orange and blue colors indicate attention scores assigned to positive and negative instances, respectively. (c)--(f) show the highest predicted probabilities for candidate labels (green) and non-candidate labels (blue) by EliMipl or MiplMa in the CRC-MIPL-Row dataset. (c) and (e) correspond to the same training bag, while (d) and (f) refer to another training bag.
• Figure 2: The MiplMa framework processes an input comprising the multi-instance bag $\boldsymbol{X}_i = \{\boldsymbol{x}_{i,1}, \boldsymbol{x}_{i,2}, \cdots, \boldsymbol{x}_{i,9}\}$ and the candidate label set $\mathcal{S}_i = \{2,3,5,7\}$, where $\mathcal{L}_d$ and $\mathcal{L}_m$ represent the dynamic disambiguation loss and the margin distribution loss, respectively.
• Figure 3: The classification accuracies (mean and std) of MiplMa with the three variants on the SIVAL-MIPL dataset ($r \in \{1,2,3\}$).
• Figure A1: The classification accuracies (mean and std) of MiplMa, M3pl, and Pl-svm.
• Figure A2: Attention scores of MiplMa, EliMipl, and DeMipl for three test bags in the FMNIST-MIPL dataset with $r=1$. The horizontal axis denotes the indices of instances, while the vertical axis represents the corresponding attention scores. Orange and blue colors indicate attention scores assigned to positive and negative instances, respectively.

Theorems & Definitions (5)

• Theorem 1
• Definition 1
• Definition 2
• Theorem 1
• proof