LAIM discretization for multi-label data

Multi-label learning is a challenging task in data mining which has attracted growing attention in recent years. Despite the fact that many multi-label datasets have continuous features, general algorithms developed specially to transform multi-label datasets with continuous attributes’ values into a finite number of intervals have not been proposed to date. Many classification algorithms require discrete values as the input and studies have shown that supervised discretization may improve classification performance. This paper presents a Label-Attribute Interdependence Maximization (LAIM) discretization method for multi-label data. LAIM is inspired in the discretization heuristic of CAIM for single-label classification. The maximization of the label-attribute interdependence is expected to improve labels prediction in data separated through disjoint intervals. The main aim of this paper is to present a discretization method specifically designed to deal with multi-label data and to analyze whether this can improve the performance of multi-label learning methods. To this end, the experimental analysis evaluates the performance of 12 multi-label learning algorithms (transformation, adaptation, and ensemble-based) on a series of 16 multi-label datasets with and without supervised and unsupervised discretization, showing that LAIM discretization improves the performance for many algorithms and measures.