HyperGraph based Imbalance Multi-label Classification with Label Specific Features

The multi-label classification problem involves assigning multiple labels to each instance, which can belong to one or more classes simultaneously. Binary relevance (BR) is a well-known paradigm for multi-label classification but it has a number of drawbacks such as class-imbalance, label inconsistency, and label correlations. While label correlations are often examined in pairwise relationships between instances, real-world systems are often better represented as networks that model complex interactions. In such cases, hypergraphs are more suitable than simple Laplacians. Our proposed approach addresses these issues by applying different weights to positive and negative examples based on class distribution to handle class imbalance, uses label-specific features to tackle inconsistency, and employing hypergraphs to capture complex sample associations. We utilize hinge loss function to reduce sensitivity to outliers and the Accelerated Proximal Gradient (APG) method to efficiently solve the optimization problem. Our approach shows competitive performance compared to existing state-of-the-art methods.