Supervised classification is the most common task in machine learning. It is included in many applications, including spam detection, face recognition, and sentiment analysis. This paper investigates the potential application of harmony search, an evolutionary meta-heuristic inspired by music improvisation, to facilitate this kind of task.
Based on this idea, the authors build three algorithms. The first classifier, a harmony-based classifier (HC), is dedicated to batch data. Incremental HC (IHC) deals with data streams, and improved IHC (IIHC) takes the well-known phenomenon of concept drift into account. The performance of HC is evaluated on a set of well-known datasets used in classification, and is compared with various classifiers such as naïve Bayes, decision tree, and neuronal network (but not powerful support vector machines). The experimental methodology is not fully detailed, so it is not clear whether the authors correctly estimate the error using held-out data.
IIHC is positively compared on three data stream sets with the ambiguous concept-adapting very fast decision tree (aCVFDT), another recent algorithm that handles concept drift. IIHC is mostly derived from StreamMiner, a classifier based on ensembles built with decision trees. The authors claim that their algorithm solves its limitations, particularly the computational overhead. Unfortunately, they do not explicitly compare IIHC to StreamMiner.