A hierarchical approach to classification for systems with complex low-level interactions

Learning in multiple steps or layers is useful when the system under study is characterized by the complex interaction of low level components. In this case it is convenient to decompose the classification problem into different layers of complexity, starting at the bottom with all low-level features, and progressing to the top through the construction of more abstract terms. In this paper we propose a hierarchical approach to classification where each layer is an attempt to improve the predictive accuracy of our classifier through the construction of new terms. We perform an experimental study of this algorithm in eighteen real-world domains; a comparison with decision trees denotes an advantage in predictive accuracy, especially when the complexity of the domain requires the construction of multiple hierarchical layers.