Option Predictive Clustering Trees for Multi-label Classification
Tomaž Stepišnik, Dragi Kocev, Sašo Džeroski · Acta Polytechnica Hungarica · 2020
In this work, we focus on the task of multi-label classification (MLC), where every example is associated with a set of labels.We present an algorithm for learning option predictive clustering trees (OPCTs) for MLC, based on the predictive clustering framework.The algorithm addresses the myopia of the standard tree induction algorithm by considering alternative splits in the internal nodes of the tree and introducing option nodes where appropriate.An option tree can be viewed as a compact representation of an ensemble, as well as, used as a pool of candidates from which a single tree can be extracted.This broadens the space of trees that is searched and reduces the myopia, compared to the standard tree induction.We evaluate the proposed OPCTs on 12 benchmark MLC datasets from different domains.Results show that OPCTs as ensembles can achieve performance similar to the bagging ensembles of PCTs, while the single trees extracted from OPCTs can outperform standard PCTs.We also perform parameter sensitivity analysis and provide avenues for future work.