[Paper Review] GM-PLL: Graph Matching based Partial Label Learning

Gengyu Lyu, Songhe Feng+3 more|arXiv (Cornell University)|Jan 10, 2019

Text and Document Classification Technologies37 references4 citations

TL;DR

This paper proposes GM-PLL, a novel graph matching-based framework for Partial Label Learning that reformulates the task as an instance-to-label matching selection problem. By extending probabilistic graph matching to a many-to-one constraint and introducing a relaxed prediction model, GM-PLL effectively disambiguates candidate labels and achieves state-of-the-art or competitive performance on both synthetic and real-world datasets.

ABSTRACT

Partial Label Learning (PLL) aims to learn from the data where each training example is associated with a set of candidate labels, among which only one is correct. The key to deal with such problem is to disambiguate the candidate label sets and obtain the correct assignments between instances and their candidate labels. In this paper, we interpret such assignments as instance-to-label matchings, and reformulate the task of PLL as a matching selection problem. To model such problem, we propose a novel Graph Matching based Partial Label Learning (GM-PLL) framework, where Graph Matching (GM) scheme is incorporated owing to its excellent capability of exploiting the instance and label relationship. Meanwhile, since conventional one-to-one GM algorithm does not satisfy the constraint of PLL problem that multiple instances may correspond to the same label, we extend a traditional one-to-one probabilistic matching algorithm to the many-to-one constraint, and make the proposed framework accommodate to the PLL problem. Moreover, we also propose a relaxed matching prediction model, which can improve the prediction accuracy via GM strategy. Extensive experiments on both artificial and real-world data sets demonstrate that the proposed method can achieve superior or comparable performance against the state-of-the-art methods.

Motivation & Objective

• Address the challenge of learning from training data where each instance is associated with a set of candidate labels, only one of which is correct.
• Overcome limitations in existing Partial Label Learning (PLL) methods that either ignore global instance relationships or fail to explicitly model instance-label assignment co-occurrence.
• Reformulate PLL as a structured matching selection problem to jointly model instance relationships and optimal instance-to-label assignments.
• Develop a graph matching framework that accommodates the many-to-one constraint inherent in PLL, where multiple instances can map to the same label.
• Improve prediction accuracy through a relaxed graph matching prediction model that enhances robustness and generalization.

Proposed method

• Reframe Partial Label Learning as an instance-to-label matching selection problem, treating correct label assignments as matchings to be identified.
• Propose a many-to-one constrained graph matching algorithm by extending a traditional one-to-one probabilistic graph matching scheme to handle multiple instances mapping to a single label.
• Incorporate both instance-level similarity and label co-occurrence patterns into a unified graph structure to model relationships across data.
• Introduce a relaxed graph matching prediction model that allows for uncertainty in label assignments, improving robustness against noisy candidate labels.
• Utilize a threshold parameter β to control the trade-off between preserving useful information and filtering out noise during matching.
• Determine the number of candidate labels r for unseen instances using a formula based on total and average label counts, validated via cross-validation.

Experimental results

Research questions

• RQ1 Can graph matching be effectively adapted to the partial label learning problem by modeling instance-to-label assignments as matchings?
• RQ2 How can the many-to-one constraint of PLL be formally integrated into a probabilistic graph matching framework?
• RQ3 Does incorporating both instance relationships and label co-occurrence patterns improve disambiguation and prediction accuracy in PLL?
• RQ4 Can a relaxed graph matching prediction model enhance robustness and performance compared to standard matching approaches?
• RQ5 How do hyperparameters like β and r affect the model’s generalization and performance across diverse datasets?

Key findings

• GM-PLL achieves superior or competitive performance compared to state-of-the-art PLL methods on both artificial and real-world datasets.
• The optimal threshold β for noise filtering was determined via cross-validation, with values ranging from 0.3 to 0.8 across datasets, indicating dataset-specific sensitivity.
• The number of candidate labels r for unseen instances was empirically validated to follow a predictable trend, with r* calculated via Eq (10) closely matching the empirically optimal r** on most datasets.
• On most UCI datasets, the total running time of GM-PLL was under 30 seconds, and on all real-world datasets, it remained under 1.5 hours, demonstrating acceptable computational efficiency.
• The testing time for GM-PLL was minimal, with the longest inference time being 422 seconds on the SoccerPlayer dataset, indicating scalability for inference.
• The total running time of GM-PLL was only 1/10 of that of PL-LEAF on all real-world datasets, showing significant efficiency improvements over some prior methods.