Identifying hidden contexts in classification

In this study we investigate how to identify hidden contexts from the data in classification tasks. Contexts are artifacts in the data, which do not predict the class label directly. For instance, in speech recognition task speakers might have different accents, which do not directly discriminate between the spoken words. Identifying hidden contexts is considered as data preprocessing task, which can help to build more accurate classifiers, tailored for particular contexts and give an insight into the data structure. We present three techniques to identify hidden contexts, which hide class label information from the input data and partition it using clustering techniques. We form a collection of performance measures to ensure that the resulting contexts are valid. We evaluate the performance of the proposed techniques on thirty real datasets. We present a case study illustrating how the identified contexts can be used to build specialized more accurate classifiers.