Cascade Classifiers for Hierarchical Decision Systems

Hierarchical classifiers are usually defined as methods of classifying inputs into defined output categories. The classification occurs first on a low-level with highly specific pieces of input data. The classifications of the individual pieces of data are then combined systematically and classified on a higher level iteratively until one output is produced. This final output is the overall classification of the data. In this paper we follow a controlled devise type of approach. The initial group of classifiers is trained using all objects in an information system S partitioned by values of the decision attribute d at its all granularity levels (one classifier per level). Only values of the highest granularity level (corresponding granules are the largest) are used to split S into information sub-systems where each one is built by selecting objects in S of the same decision value. These sub-systems are used for training new classifiers at all granularity levels of its decision attribute. Next, we split each sub-system further by sub-values of its decision value. The obtained tree-structure with groups of classifiers assigned to each of its nodes is called a cascade classifier. Given an incomplete information system with a hierarchical decision attribute d, we consider the problem of training classifiers describing values of d at its lowest granularity level. Taking MIRAI database of music instrument sounds (Raś et al., 2007b), as an example, we show that the confidence of such classifiers can be lower than the confidence of cascade classifiers.