Mitigating Instance Entanglement in Instance-Dependent Partial Label Learning

Rui Zhao; Bin Shi; Kai Sun; Bo Dong

cs.CVMar 5, 2026

Mitigating Instance Entanglement in Instance-Dependent Partial Label Learning

Authors: Rui Zhao, Bin Shi, Kai Sun, Bo Dong

Organizations: School of Computer Science and Technology, Xi’an Jiaotong University · Shaanxi Province Key Laboratory of Big Data Knowledge Engineering, Xi’an Jiaotong University · School of Distance Education, Xi’an Jiaotong University

Abstract

Partial label learning is a prominent weakly supervised classification task, where each training instance is ambiguously labeled with a set of candidate labels. In real-world scenarios, candidate labels are often influenced by instance features, leading to the emergence of instance-dependent PLL (ID-PLL), a setting that more accurately reflects this relationship. A significant challenge in ID-PLL is instance entanglement, where instances from similar classes share overlapping features and candidate labels, resulting in increased class confusion. To address this issue, we propose a novel Class-specific Augmentation based Disentanglement (CAD) framework, which tackles instance entanglement by both intra- and inter-class regulations. For intra-class regulation, CAD amplifies class-specific features to generate class-wise augmentations and aligns same-class augmentations across instances. For inter-class regulation, CAD introduces a weighted penalty loss function that applies stronger penalties to more ambiguous labels, encouraging larger inter-class distances. By jointly applying intra- and inter-class regulations, CAD improves the clarity of class boundaries and reduces class confusion caused by entanglement. Extensive experimental results demonstrate the effectiveness of CAD in mitigating the entanglement problem and enhancing ID-PLL performance. The code is available at https://github.com/RyanZhaoIc/CAD.git.

Figures & tables

Figure 1: Illustration of how solely reducing intra-class distance may lead to class confusion (Figure (a)) and how class-specific augmentation combined with confidence-based penalties can mitigate confusion by increasing inter-class distance (Figure (b)). In Figure (a), similar instances with shared candidate labels may be misaligned. In Figure (b), class-specific augmentations are aligned while simultaneously lowering the confidence of confusing labels.

Figure 2: Illustration of CAD framework. The upper part forces the model to produce outputs with less ambiguity based on weighted loss. The lower part generates class-specific augmentations that amplify specific features and guide them to have aligned representations. “//” denotes where gradient propagation is stopped. In the figure, “Q1-Dog” and “K1-Dog” represent the query and key embeddings for dog class-specific augmentations of img1, respectively. Here, augmentations guided by the same label are represented using the same color.

$\xi$	0.90		0.95		0.99
	#Pairs	#Instances	#Pairs	#Instances	#Pairs	#Instances
Fashion-MNIST	528,996	38,717	125,535	16,992	3,658	1,802
CIFAR-10	54,531	22,455	3,101	2,399	2	4
CIFAR-100	1,991	1,482	74	108	17	34

Table 1: Statistics of entangled instances under different similarity thresholds. “#Pairs” and “#Instances” denote the number of entangled pairs and the number of unique instances involved, respectively. Note that one instance may appear in multiple pairs, so the counts are not in a one-to-one correspondence.

Figure 3: Training accuracy on entangled instances under different similarity thresholds. “0” indicates that all entangled instances are used, reflecting the original accuracy. All models are trained on the full training set, and the entangled subsets are used for evaluation.

	Fashion-MNIST	CIFAR-10	CIFAR-100	Flower	Oxford-IIIT Pet
POP	81.91 ± 0.37	89.55 ± 0.36	64.57 ± 0.37	33.02 ± 2.62	53.83 ± 1.55
VALEN	82.91 ± 0.12	81.29 ± 0.39	60.19 ± 0.82	32.89 ± 0.88	50.74 ± 0.91
IDGP	85.14 ± 0.39	84.12 ± 0.99	62.27 ± 1.89	41.26 ± 1.33	59.74 ± 0.71
CP-DPLL	79.67 ± 6.95	80.29 ± 0.18	56.93 ± 1.47	42.42 ± 0.12	49.13 ± 2.49
CAVL	78.06 ± 3.10	79.89 ± 0.07	59.76 ± 0.49	40.69 ± 1.93	50.64 ± 0.11
PICO	85.92 ± 0.38	82.91 ± 1.43	58.56 ± 0.13	33.76 ± 1.30	62.64 ± 0.45

Table 2: Average accuracy and standard deviation (%) over 5 runs. Best results are highlighted in bold , and the second best are underlined .

		0.100%		0.010%		0.001%
		Acc	Distance	Acc	Distance	Acc	Distance
F-MNIST	RC	83.38	0.4507	74.02	0.4036	68.75	0.3623
	ABLE	87.75	0.0955	78.89	0.0798	72.32	0.0865
	DIRK	89.57	0.4890	80.91	0.4110	72.86	0.3447
	CAD	89.87	0.5099	81.39	0.4372	74.79	0.4186
CIFAR-10	RC	84.51	0.6161	79.17	0.6088	70.42	0.5588

Table 3: Accuracy on entangled instances and average Euclidean distance between entangled pairs.

Panel of Figure 4 — Figure 4: t-SNE visualization on the CIFAR-10 benchmark.

Panel of Figure 5 — Figure 5: Confusion matrices on the CIFAR-10 benchmark.

Metric	RC	ABLE	DIRK	CAD
Instance	0.040	0.007	0.035	0.057
Avg. Pairwise	0.414	0.128	0.333	0.337
Centroid	1.009	0.340	0.936	1.103

Table 4: Comparison of distances between similar classes on the CIFAR-10 benchmark (higher is better).

Dataset	CAD	w/o CA	w/o RL	w/o Both
Fashion-MNIST	92.14 $\pm$ 0.91	91.19 $\pm$ 0.22	91.48 $\pm$ 0.61	85.30 $\pm$ 0.87
CIFAR-10	93.57 $\pm$ 0.24	93.32 $\pm$ 0.13	91.21 $\pm$ 0.52	87.81 $\pm$ 1.13

Table 5: Ablation accuracy (%, mean

\pm

std).

Method	Accuracy (%)
CAD-CAM	74.56 ± 3.16
CAD w/o cues	69.46 ± 2.35
CAD with fine-grained prompting	76.23 ± 2.17

Table 6: Accuracy (%) with different prompting.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Train	Test	Dims	Classes	Labels
Fashion-MNIST	60,000	10,000	$28\times 28$	10	7.40
CIFAR-10	50,000	10,000	$32\times 32$	10	5.89
CIFAR-100	50,000	10,000	$32\times 32$	100	9.40
Flower	1,020	6,149	$224\times 224$	102	5.49
Oxford-IIIT Pet	3,680	3,669	$224\times 224$	37	3.73

Appendix

Table 1: The statistics of datasets. Here, labels refers to the average number of candidate labels assigned to each instance.

Ratio	0.100%			0.010%			0.001%
	#Pairs	#instances	$\xi$	#Pairs	#instances	$\xi$	#Pairs	#instances	$\xi$
Fashion-MNIST	822,911	42,851	0.8774	82,291	13,345	0.9594	8,229	3,114	0.9859
CIFAR-10	374,916	47,655	0.8352	37,491	17,513	0.9088	3,749	2,833	0.9478
CIFAR-100	23,041	10,007	0.8303	2,304	1,644	0.8,971	230	270	0.9364
Oxford-IIIT	133	131	0.8972	13	22	0.9211	1	2	0.9407
Flower	5	8	0.7740	1	2	0.8137	1	2	0.8137

Appendix

Table 2: Statistic of entangled instances at different thresholds. #Pairs and #Instances represent the number of entangled pairs and instances, respectively. Since each instance can be entangled with multiple other instances, the number of pairs and instances may not follow a strict ratio.

Method	CIFAR-10	CIFAR-100
DIRK	10.19	14.37
ABLE	13.89	19.15
CAD-CAM	18.50	25.23
CAD	24.66	35.54

Appendix

Table 3: GPU-hours on CIFAR-10 and CIFAR-100 benchmarks.

	Fashion-MNIST	CIFAR-10	CIFAR-100
RC-C	66.66 ± 0.99	76.80 ± 0.27	59.63 ± 0.20
RC	84.87 ± 1.48	87.53 ± 0.94	65.26 ± 0.40
ABLE-C	78.41 ± 0.35	59.66 ± 1.04	58.03 ± 0.28
ABLE	89.81 ± 0.08	83.92 ± 0.67	63.92 ± 0.39
DIRK-C	79.05 ± 0.49	85.06 ± 0.15	62.66 ± 0.18
DIRK	91.48 ± 0.21	91.48 ± 0.21	68.77 ± 0.49

Appendix

Table 4: Comparison of accuracy (expressed in percentage) with Diffusion-based Augmentation. The best results for each benchmark are highlighted in bold .

Method	CIFAR-10	CIFAR-100
POP	89.55 ± 0.36	64.57 ± 0.37
POP+CAD	91.56 ± 0.80	66.56 ± 2.86
ABLE	83.92 ± 0.67	63.92 ± 0.39
ABLE+CAD	85.21 ± 4.08	68.63 ± 1.13
DIRK	90.87 ± 0.25	68.77 ± 0.49
DIRK+CAD	92.86 ± 0.17	72.35 ± 0.53

Appendix

Table 5: Accuracy (%) comparison on CIFAR-10 and CIFAR-100. “+CAD” denotes equipping the baseline with our proposed CAD contrastive learning module.

Method	Accuracy (%)
ABLE	91.82
DIRK	92.91
CAD (Ours)	93.24

Appendix

Table 6: Accuracy (%) on the PLCIFAR10 dataset with partial labels from human annotations.

Panel of Figure 1 — Figure 1: T-SNE visualization on the Fashion-MNIST benchmark, with different colors representing different classes.

Panel of Figure 2 — Figure 2: Confusion Matrix on the Fashion-MNIST benchmark.

Panel of Figure 3 — Figure 3: Ablation of Confusion Matrix on the Fashion-MNIST benchmark.

$\beta$	0.00	0.25	0.50	0.75	1.00	1.25	1.50	1.75	2.00
Accuracy	89.19	92.53	92.95	92.94	93.85	94.10	93.97	93.88	92.87

Appendix

Table 7: Accuracy with different

\beta

in the CIFAR-10 benchmark. The best results are highlighted in bold .

Method	AGNews Accuracy (%)
DIRK	77.34
ABLE	93.92
CAD (Ours)	94.30

Appendix

Table 8: Text classification accuracy (%) on AGNews. All methods use BERT-base as the student encoder; augmented data are generated using T5-base with class-specific prompts.

$\xi$	F-MNIST	CIFAR-10	CIFAR-100
0.90	16.01	27.45	79.43
0.95	15.28	26.56	51.06
0.99	11.51	25.00	38.46

Appendix

Table 9: Recovery rate (%) of entangled instances misclassified by DIRK but correctly classified under full supervision (ground-truth labels), at different similarity thresholds

\xi

Figure 5: Diffusion-based visualization of the class-specific augmentations generated from CIFAR-10. The left side shows the samples with their ground truth labels, while the right side displays the augmentations generated with different candidate label guidance.

Method	Fashion-MNIST	CIFAR-10	CIFAR-100	Flower	Oxford-IIIT Pet
Original ( $x$ )	-0.0276 $\mid$ 0.2364	-0.0415 $\mid$ 0.2151	-0.0416 $\mid$ 0.2176	-0.0010 $\mid$ 0.2185	-0.0002 $\mid$ 0.2282
Diffusion ( $x^{\prime}$ )	0.0120 $\mid$ 0.2555	0.0314 $\mid$ 0.2566	0.0181 $\mid$ 0.2453	0.0243 $\mid$ 0.2784	0.0169 $\mid$ 0.2705

Appendix

Table 10: Comparison of CLIP-score margin and target similarity (margin

\mid

CLIP-score).

Figure 6: CAM-based visualization of the class-specific augmentations generated from CIFAR-10. The left side shows the samples with their ground truth labels, while the right side displays the augmentations generated with different candidate label guidance.

Weakly Supervised Learning

Weakly supervised learning trains models from incomplete, noisy, coarse, or indirect labels instead of dense ground truth.

See all 156 papers

Calibratable Disambiguation Loss for Multi-Instance Partial-Label Learning
Dec 19, 2025Wei Tang, Yin-Fang Yang, Weijia Zhang +1Model Calibration Weakly Supervised Learning
School of Computer Science and Engineering, Southeast University, Nanjing 210096, China, and the Key Laboratory of Computer Network and Information Integration (Southeast University), MoE, China · School of Computer and Information Sciences, The University of Newcastle, Callaghan, NSW 2308, Australia
Feature-Label Modal Alignment for Robust Partial Multi-Label Learning
Apr 10, 2026Yu Chen, Weijun Lv, Yue Huang +4Multi-Label Classification Learning with Noisy Labels
School of Automation, Guangdong University of Technology, Guangzhou 510006, China · Shenzhen Key Laboratory of Visual Object Detection and Recognition, Harbin Institute of Technology, Shenzhen 518055, China · School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou 510275, China
Bringing Clustering to MLL: Weakly-Supervised Clustering for Partial Multi-Label Learning
Apr 10, 2026Yu Chen, Weijun Lv, Yue Huang +2Multi-Label Classification Learning with Noisy Labels
School of Automation, Guangdong University of Technology

Dec 19, 2025J/K move · Enter open · S save

Calibratable Disambiguation Loss for Multi-Instance Partial-Label Learning

Wei Tang, Yin-Fang Yang, Weijia Zhang, Min-Ling Zhang

School of Computer Science and Engineering, Southeast University, Nanjing 210096, China, and the Key Laboratory of Computer Network and Information Integration (Southeast University), MoE, China · School of Computer and Information Sciences, The University of Newcastle, Callaghan, NSW 2308, Australia

Model Calibration Weakly Supervised Learning Multiple Instance Learning

Open paper AskPDF arXiv

Multi-instance partial-label learning (MIPL) is a weakly supervised framework that extends the principles of multi-instance learning (MIL) and partial-label learning (PLL) to address the challenges of inexact supervision in both instance and label spaces. However, existing MIPL approaches often suffer from poor calibration, undermining classifier reliability. In this work, we propose a plug-and-play calibratable disambiguation loss (CDL) for classification and calibration, which modulates a disambiguation objective by a top-vs-competitor prediction margin. The competitor is instantiated either as the second strongest candidate label or as the strongest non-candidate label, yielding two variants that respectively emphasize candidate-level separation and candidate-vs-non-candidate suppression. Theoretically, we analyze CDL as a margin-modulated momentum-based disambiguation loss (MDL) objective, derive a lower-bound and a pseudo-label confidence-alignment bound for calibration, and show through gradient and momentum analyses how margin shaping affects weight updates. Experimental results on benchmark and real-world MIPL datasets, together with representative PLL adaptation, confirm that our CDL significantly improves both classification accuracy and expected calibration error.

Abstract

Weakly Supervised Learning

Explore similar work

Calibratable Disambiguation Loss for Multi-Instance Partial-Label Learning

Feature-Label Modal Alignment for Robust Partial Multi-Label Learning

Bringing Clustering to MLL: Weakly-Supervised Clustering for Partial Multi-Label Learning