On Quality of Quality Measures for Classification

In this paper we provide a theoretical discussion of the impact of uncertainty in quality measurement on the expected benefits of including biometric signal quality measures in classification. While an ideal signal quality measure should be a precise quantification of the actual signal properties relevant to the classification process, a real quality measurement may be uncertain. We show how does the degree of uncertainty in quality measurement impact the gains in class separation achieved thanks to using quality measures as conditionally relevant classification feature. We demonstrate that while noisy quality measures become irrelevant classification features, they do not impair class separation beyond the baseline result. We present supporting experimental results using synthetic data.