Dynamic Prototype Weighting for Multi-label Classification Problems
Ali Shahsavandi, Mohammad Taheri, Koorush Ziarati · 2024
Multi-label classification involves the assignment of several labels to each data instance simultaneously. Binary relevance (BR) is widely acknowledged as the most straightforward approach for handling multi-label scenarios. The process involves breaking down the multi-label learning task into a number of independent binary learning tasks. Using Nearest Neighbor (NN) as the binary classifier in BR (BRNN) is a simple, descriptive, and powerful approach. However, the NN has drawbacks, including its dependence on the distance criterion. To address this, we propose a method that tunes a parametric distance using a prototype weighting approach, minimizing complete cross-validation (CCV) classification error on training data. This improves the classifier’s generalization and reduces randomness. Nevertheless, it is well established that when dealing with imbalanced data, other performance evaluation metrics, such as the F-measure, offer more relevant insights than classification error. The second contribution of this paper is extending this method to improve the F-measure. Due to the imbalanced nature of multi-label data, it is expected that performance will improve. Our method is compared to state-of-the-art NN-based approaches, demonstrating comparable or superior performance in experiments.