Label Distribution Learning by Partitioning Label Distribution Manifold.

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Label distribution learning (LDL) faces challenges with large output spaces.
• Existing methods often use clustering independent of label correlation, limiting performance.
• Local label correlation exploitation has shown promise but requires better integration.

Purpose of the Study:

• To propose a novel LDL method, LDL-PLDM, that jointly partitions data and learns label distribution manifolds.
• To improve the modeling of label correlation for enhanced LDL performance.
• To address the limitations of clustering-based data partitioning in LDL.

Main Methods:

• LDL-PLDM jointly bipartitions the training set and learns a label distribution manifold.
• The method recursively refines partitions until reconstruction error is minimized.
• This approach ensures partitions are intrinsically linked to label correlation.

Main Results:

• LDL-PLDM achieves label-correlation-aware data partitioning.
• The learned label distribution manifold effectively captures complex label correlations.
• Experimental results demonstrate statistically significant improvements over state-of-the-art LDL methods.

Conclusions:

• LDL-PLDM offers a superior approach to modeling label correlation in LDL.
• The joint optimization of data partitioning and manifold learning enhances model accuracy.
• This method provides a robust framework for tackling large output spaces in LDL.