Complementary to Multiple Labels: A Correlation-Aware Correction Approach.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Complementary Label Learning (CLL) is effective for multi-class problems by estimating a transition matrix.
• Existing multi-class CLL methods fail on multi-label data due to the assumption of single relevant labels per instance.
• This limitation arises because multi-labeled instances possess multiple co-existing relevant labels, distorting transition matrix estimation.

Purpose of the Study:

• To address the limitations of current CLL techniques in handling multi-label data.
• To theoretically analyze and demonstrate the distortion of transition matrices in multi-label scenarios.
• To propose a novel method for accurate transition matrix estimation in Multi-Label CLL (ML-CLL).

Main Methods:

• A two-step approach is proposed to estimate the transition matrix from candidate labels in ML-CLL.
• The method first decomposes the multi-label problem into binary classification tasks to estimate an initial transition matrix.
• This initial matrix is then refined using label correlations to incorporate relationships among labels, and an MSE-based regularizer is introduced.

Main Results:

• Theoretical analysis reveals distortions in transition matrix estimation for multi-label CLL when co-existing labels are ignored.
• The proposed two-step method effectively estimates the transition matrix for ML-CLL, even without pre-existing multi-labeled data.
• The method is shown to be classifier-consistent, and the MSE regularizer mitigates overfitting to noise.

Conclusions:

• The proposed method offers a robust solution for Complementary Label Learning on multi-label datasets.
• Accurate transition matrix estimation is crucial for effective ML-CLL.
• The approach enhances the applicability of CLL to more complex, real-world multi-label classification tasks.