Gradient-Based Label Binning in Multi-label Classification

In multi-label classification, where a single example may be associated with several class labels at the same time, ability to model dependencies between is considered crucial effectively optimize non-decomposable evaluation measures, such as Subset 0/1 loss. The gradient boosting framework provides well-studied foundation for learning models that are specifically tailored loss function and recent research attests achieve high predictive accuracy in setting. utilization of second-order derivatives, used by many approaches, helps guide minimization losses, due information about pairs it incorporates into optimization process. On downside, this comes computational costs, even if number small. work, we address bottleneck approach—the need solve system linear equations—by integrating novel approximation technique procedure. Based on derivatives computed during training, dynamically group predefined bins impose an upper bound dimensionality system. Our experiments, using existing rule-based algorithm, suggest boost speed without any significant performance.