ml-lab-sau / discriminatory-label-specific-weights-for-multi-label-learning-with-missing-labels

To deal with the class imbalance problem in multi-label learning with missing labels, we propose Class Imbalance aware Missing labels Multi-label Learning, CIMML. Our proposed method handles class imbalance issue by constructing a label weight matrix with weight estimation guided by how frequently a label is present, absent, and unobserved.

CIMML Model for Multi Label Leanring

➡️ Implementation of CIMML Model

➡️ If you use CIMML's Model and functionality in a scientific publication, please cite the following paper:

Cite this paper

R. Rastogi and K. Sanjay,"Discriminatory Label-specific Weights for Multi-label Learning with Missing Labels." Neural Processing Letters 55, 1397–1431 (2023).

BibTeX entry:

@Article{Rastogi2023,
author={Rastogi, Reshma and Kumar, Sanjay},
title={Discriminatory Label-specific Weights for Multi-label Learning with Missing Labels},
journal={Neural Processing Letters},
year={2023},
month={Apr},
day={01},
volume={55},
number={2},
pages={1397-1431},
abstract={Class labels in multi-label datasets are only associated with a very small fraction of the data instances leading to a class imbalance problem. There exist multi-label learning algorithms that handle the datasets' class imbalance issue by assigning favorable weights for positive labels to compensate for their scarce occurrence. The weight assignment assumes full label information and doesn't provision for the uncertainty introduced by unobserved or missing labels. Therefore, to deal with the class imbalance problem in multi-label learning with missing labels, we propose Class Imbalance aware Missing labels Multi-label Learning, CIMML. Our proposed method handles class imbalance issue by constructing a label weight matrix with weight estimation guided by how frequently a label is present, absent, and unobserved. Utilizing the class imbalance sensitive weights and auxiliary label correlations, we introduce a weighted squared loss function with discriminatory label weights which guides the missing label completion and learns the multi-label classifier efficiently. Extensive experiments on several benchmark datasets validate the effectiveness of proposed classifier against state-of-the-art multi-label learning approaches for missing labels underlining the significance of handling label imbalance for multi-label learning for missing labels.},
issn={1573-773X},
doi={10.1007/s11063-022-10945-z},
url={https://doi.org/10.1007/s11063-022-10945-z}
}