A Scalable Curiosity-Driven Game-Theoretic Framework for Long-Tail Multi-Label Learning in Data Mining

Jing Yang; Keze Wang

cs.LGFeb 17, 2026

A Scalable Curiosity-Driven Game-Theoretic Framework for Long-Tail Multi-Label Learning in Data Mining

Authors: Jing Yang, Keze Wang

Organizations: Sun Yat-sen University

Abstract

The long-tail distribution, where a few head labels dominate while rare tail labels abound, poses a persistent challenge for large-scale Multi-Label Classification (MLC) in real-world data mining applications. Existing resampling and reweighting strategies often disrupt inter-label dependencies or require brittle hyperparameter tuning, especially as the label space expands to tens of thousands of labels. To address this issue, we propose Curiosity-Driven Game-Theoretic Multi-Label Learning (CD-GTMLL), a scalable cooperative framework that recasts long-tail MLC as a multi-player game - each sub-predictor ("player") specializes in a partition of the label space, collaborating to maximize global accuracy while pursuing intrinsic curiosity rewards based on tail label rarity and inter-player disagreement. This mechanism adaptively injects learning signals into under-represented tail labels without manual balancing or tuning. We further provide a theoretical analysis showing that our CD-GTMLL converges to a tail-aware equilibrium and formally links the optimization dynamics to improvements in the Rare-F1 metric. Extensive experiments across 7 benchmarks, including extreme multi-label classification datasets with 30,000+ labels, demonstrate that CD-GTMLL consistently surpasses state-of-the-art methods, with gains up to +1.6% P@3 on Wiki10-31K. Ablation studies further confirm the contributions of both game-theoretic cooperation and curiosity-driven exploration to robust tail performance. By integrating game theory with curiosity mechanisms, CD-GTMLL not only enhances model efficiency in resource-constrained environments but also paves the way for more adaptive learning in imbalanced data scenarios across industries like e-commerce and healthcare.

Figures & tables

Figure 1. Precision@k (mean ± std, %) results on three benchmark datasets.

Figure 2. Rare-F1 (mean ± std, %) results on artificially down-sampled “-R” variants. @X% indicates the fraction of positives removed.

Method	Pascal VOC		COCO		Yeast		Mediamill
	mAP	Rare-F1	mAP	Rare-F1	Micro-F1	Rare-F1	Macro-F1	Rare-F1
CC ( Read et al., 2009 )	$89.8\pm 0.4$	$73.6\pm 0.5$	$63.9\pm 0.5$	$42.7\pm 0.6$	$76.8\pm 0.4$	$66.4\pm 0.5$	$50.7\pm 0.6$	$37.1\pm 0.7$
ML-GCN ( Chen et al., 2019 )	$91.0\pm 0.3$	$76.0\pm 0.4$	$66.1\pm 0.4$	$45.8\pm 0.5$	$78.1\pm 0.3$	$67.6\pm 0.4$	$53.3\pm 0.5$	$39.3\pm 0.6$
C-Tran ( Lanchantin et al., 2021 )	$91.3\pm 0.3$	$76.4\pm 0.4$	$66.3\pm 0.4$	$46.1\pm 0.5$	$78.2\pm 0.3$	$67.7\pm 0.4$	$54.0\pm 0.5$	$39.9\pm 0.6$
ML-Decoder ( Ridnik et al., 2023 )	$91.9\pm 0.3$	$77.6\pm 0.4$	$66.9\pm 0.4$	$46.8\pm 0.5$	$79.0\pm 0.3$	$68.3\pm 0.4$	$55.1\pm 0.4$	$40.9\pm 0.5$
LCIFS ( Fan et al., 2024 )	$91.4\pm 0.3$	$77.8\pm 0.4$	$67.1\pm 0.4$	$47.3\pm 0.5$	$78.5\pm 0.3$	$67.6\pm 0.4$	$55.4\pm 0.4$	$41.0\pm 0.5$

Table 1. Mean performance results (mean ± std, %) on standard-frequency datasets. Rare-F1 is computed on the bottom 20% of labels. Best and second-best results are in bold and underlined , respectively.

Dataset	Model Variant	mAP (%)	Micro-F1 (%)	Rare-F1 (%)
Pascal VOC 2007	Full CD-GTMLL	92.6 $\pm$ 0.2	91.0 $\pm$ 0.3	79.2 $\pm$ 0.3
	No-Curiosity ( $\alpha=0$ )	92.3 $\pm$ 0.3	90.5 $\pm$ 0.3	75.2 $\pm$ 0.4
	Single-Predictor ( $N=1$ )	91.8 $\pm$ 0.4	90.0 $\pm$ 0.4	74.8 $\pm$ 0.5
MS-COCO 2014	Full CD-GTMLL	68.4 $\pm$ 0.3	61.5 $\pm$ 0.3	49.3 $\pm$ 0.4
	No-Curiosity ( $\alpha=0$ )	66.0 $\pm$ 0.4	60.0 $\pm$ 0.4	46.1 $\pm$ 0.5
	Single-Predictor ( $N=1$ )	65.5 $\pm$ 0.5	59.5 $\pm$ 0.5	45.9 $\pm$ 0.6

Table 2. Ablation study results (mean

\pm

std, %) on Pascal VOC 2007 and MS-COCO 2014.

Dataset	Label Type	Player 1	Player 2	Player 3
Pascal VOC 2007	Head (Frequent)	$2.1\pm 0.3$	$2.4\pm 0.2$	$1.5\pm 0.2$
Pascal VOC 2007	Tail (Rare)	$2.5\pm 0.3$	$1.4\pm 0.2$	$2.1\pm 0.3$
MS-COCO 2014	Head (Frequent)	$1.4\pm 0.2$	$2.3\pm 0.3$	$2.3\pm 0.2$
MS-COCO 2014	Tail (Rare)	$2.1\pm 0.4$	$1.5\pm 0.3$	$2.4\pm 0.4$

Table 3. Mean performance rank (mean ± std) of each player on different label sets for Pascal VOC and MS-COCO. A lower rank indicates better performance.

Figure 3. Mean performance rank (mean ± std) of each player on different label sets for Pascal VOC and MS-COCO. A lower rank indicates better performance.

Figure 4. Comparison of computational overhead and performance on the MS-COCO dataset.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Disagreement Weight ( $\beta$ )	mAP (%)	Rare-F1 (%)
0.0	92.0 $\pm$ 0.4	76.5 $\pm$ 0.5
0.05	92.4 $\pm$ 0.3	78.1 $\pm$ 0.4
0.1	92.5 $\pm$ 0.3	78.9 $\pm$ 0.4
0.2	92.6 $\pm$ 0.2	79.3 $\pm$ 0.3
0.5	92.4 $\pm$ 0.3	79.0 $\pm$ 0.4
1.0	92.1 $\pm$ 0.4	78.6 $\pm$ 0.4

Appendix

Table 4. Sensitivity Analyses on the disagreement weight

\beta

on Pascal VOC 2007. Results are reported as mean

\pm

std over 3 runs. Best performance is in bold.

Overlap Ratio ( $\rho$ )	mAP (%)	Rare-F1 (%)
0.0	91.5 $\pm$ 0.4	75.8 $\pm$ 0.6
0.1	92.3 $\pm$ 0.3	78.5 $\pm$ 0.4
0.2	92.6 $\pm$ 0.2	79.3 $\pm$ 0.3
0.3	92.5 $\pm$ 0.3	79.1 $\pm$ 0.3
0.4	92.4 $\pm$ 0.3	78.9 $\pm$ 0.4
0.5	92.4 $\pm$ 0.4	78.8 $\pm$ 0.4

Appendix

Table 5. Sensitivity analyses on the label overlap ratio

\rho

on Pascal VOC 2007. Results are reported as mean

\pm

std over 3 runs. Best performance is in bold.

Game-Theoretic RL

See all 28 papers

May 2, 2026

CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning

Long-tailed distributions are common in real-world recognition tasks, where a few head classes have many samples while most tail classes have very few. Recently, fine-tuning foundation models for long-tailed learning has gained attention due to their excellent performance. However, most existing methods focus solely on mitigating long-tailed distribution bias while overlooking concept confusion caused by the long-tailed distribution. In this paper, we study this problem and attribute it to the mutual exclusivity of single-label supervision under long-tailed distributions, which suppresses feature sharing among related classes and amplifies the dominance of head classes, leading to disrupted inter-class discriminability. To address this, we propose CUE, Concept-aware mUlti-label Expansion, which introduces multi-label concept signals to preserve disrupted inter-class relationships. Specifically, CUE constructs concept sets by (i) extracting instance-level visual cues from zero-shot CLIP and (ii) generating class-level semantic cues with LLM; the two cues are incorporated via separately weighted Binary Logit-Adjustment (BLA) auxiliary losses and jointly optimized with the baseline Logit-Adjustment (LA) loss. Experiments on several long-tailed benchmarks, CUE achieves balanced and strong performance, surpassing recent state-of-the-art methods. Code is available at: https://github.com/zhangruichi/CUE.

Ruichi Zhang, Chikai Shang, Jiacheng Yang +4

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China · College of Computer Science and Software Engineering, Shenzhen University, China · Institute of High Performance Computing, A*STAR, Singapore

Long-Tailed Classification Logit Adjustment Vision Foundation Model Adaptation

Ask PDF

May 27, 2026

Principled Algorithms for Optimizing Generalized Metrics in Multi-Label Learning

Many real-world classification tasks require predicting multiple labels per instance, necessitating the optimization of complex evaluation metrics such as the

F

-measure and Jaccard index. While the Empirical Utility Maximization (EUM) framework is natural for these population-level metrics, existing theoretical results are largely limited to asymptotic Bayes-consistency. In this paper, we develop principled learning algorithms for optimizing a broad class of generalized metrics within the EUM framework, grounded in the stronger notion of

H

-consistency. Our key contribution is the design of novel surrogate loss functions for multi-label learning that admit provable

H

-consistency bounds, enabling optimization with non-asymptotic guarantees tailored to the hypothesis class and finite samples. Crucially, we prove these combinatorially formulated surrogates decompose exactly, operating in strictly

O(l)

time without approximations. Building on this foundation, we introduce MMO (Multi-Label Metric Optimization), a new family of algorithms for optimizing generalized linear-fractional metrics. We validate our approach through extensive experiments, demonstrating robust scalability and superior performance over state-of-the-art continuous baselines on large-scale datasets (MS-COCO, Reuters-21578) in high-sparsity, deep learning regimes. Our results offer both theoretical rigor and practical effectiveness for general multi-label metric optimization.

Mehryar Mohri, Yutao Zhong

Google Research & CIMS · New York, NY 10011 · Google Research

Multi-Label Classification Statistical Learning Theory

Ask PDF

Mar 22, 2026

Reframing Long-Tailed Learning via Loss Landscape Geometry

Balancing performance trade-off on long-tail (LT) data distributions remains a long-standing challenge. In this paper, we posit that this dilemma stems from a phenomenon called "tail performance degradation" (the model tends to severely overfit on head classes while quickly forgetting tail classes) and pose a solution from a loss landscape perspective. We observe that different classes possess divergent convergence points in the loss landscape. Besides, this divergence is aggravated when the model settles into sharp and non-robust minima, rather than a shared and flat solution that is beneficial for all classes. In light of this, we propose a continual learning inspired framework to prevent "tail performance degradation". To avoid inefficient per-class parameter preservation, a Grouped Knowledge Preservation module is proposed to memorize group-specific convergence parameters, promoting convergence towards a shared solution. Concurrently, our framework integrates a Grouped Sharpness Aware module to seek flatter minima by explicitly addressing the geometry of the loss landscape. Notably, our framework requires neither external training samples nor pre-trained models, facilitating the broad applicability. Extensive experiments on four benchmarks demonstrate significant performance gains over state-of-the-art methods. The code is available at:https://gkp-gsa.github.io/.

Shenghan Chen, Yiming Liu, Yanzhen Wang +2

Shandong University · Zhejiang Sci-Tech University

Long-Tail Learning Loss Landscape Geometry

Ask PDF

Abstract

Game-Theoretic RL

Explore similar work

CUE: Concept-Aware Multi-Label Expansion to Mitigate Concept Confusion in Long-Tailed Learning

Principled Algorithms for Optimizing Generalized Metrics in Multi-Label Learning

Reframing Long-Tailed Learning via Loss Landscape Geometry