Evaluating ReliefF-Based Multi-Label Feature Selection Algorithm

In multi-label learning, each instance is associated with multiple labels, which are often correlated. As other machine learning tasks, multi-label learning also suffers from the curse of dimensionality, which can be mitigated by feature selection. This work experimentally evaluates four multi-label feature selection algorithms that use the filter approach. Three of them are based on the ReliefF algorithm, which takes into account interacting features. The quality of the selected features is assessed by three different learning algorithms. Evaluating multi-label learning algorithms is a complicated task, as multiple evaluation measures, which might optimize different loss functions, should be considered. To this end, \(General_B\), a baseline algorithm which learns by only looking at the multi-labels of the dataset, is used as a reference. Results show that feature selection contributed to improve the performance of classifiers initially worse than \(General_B\) and highlight ReliefF-based algorithms in some experimental settings.