Abstract
In this work, we analyze the cross-entropy function, widely used in classifiers both as a performance measure and as an optimization objective. We contextualize cross-entropy in the light of Bayesian decision theory, the formal probabilistic framework for making decisions, and we thoroughly analyze its motivation, meaning and interpretation from an information-theoretical point of view. In this sense, this article presents several contributions: First, we explicitly analyze the contribution to cross-entropy of (i) prior knowledge; and (ii) the value of the features in the form of a likelihood ratio. Second, we introduce a decomposition of cross-entropy into two components: discrimination and calibration. This decomposition enables the measurement of different performance aspects of a classifier in a more precise way; and justifies previously reported strategies to obtain reliable probabilities by means of the calibration of the output of a discriminating classifier. Third, we give different information-theoretical interpretations of cross-entropy, which can be useful in different application scenarios, and which are related to the concept of reference probabilities. Fourth, we present an analysis tool, the Empirical Cross-Entropy (ECE) plot, a compact representation of cross-entropy and its aforementioned decomposition. We show the power of ECE plots, as compared to other classical performance representations, in two diverse experimental examples: a speaker verification system, and a forensic case where some glass findings are present.
Highlights
Probabilistic approaches for data mining, machine learning and pattern recognition have proven their effectiveness both theoretically and practically in multiple applications [1]
This article presents sound contributions for the general fields of pattern recognition, data mining and machine learning. One of these contributions is showing the, typically independent, sources of information affecting cross-entropy: prior knowledge and value of the features. These are not typically taken into account by previous approaches using cross-entropy in classifiers, such as in [3,5,6], where it is most common that the empirical prior probability is used solely
Other related measures such as Confusion Entropy (CEN) or Matthews Correlation Coefficient (MCC) [41,42] work with decision errors rather than probabilities, which implies the selection of a threshold τ, and they do not consider performance at different prior probabilities either
Summary
Probabilistic approaches for data mining, machine learning and pattern recognition have proven their effectiveness both theoretically and practically in multiple applications [1]. Probabilistic outputs of systems have proven to be useful in many other research and application areas such as clinical decision support systems [9], cognitive psychology [10,11], biometric systems [12,13,14], weather forecasting [15] and forensic science [16,17] In all those areas, as well as it happens with classifiers in general, Bayesian decision theory [1] constitutes the formal framework to make optimal choices of courses of action. Apart from its advantages for general classifiers [13], this analysis is of particular interest in many applications such as forensic science [17,24], where prior probabilities and likelihood ratios are computed by different agents, with different responsibilities in the decision process. We generalize the use of an analysis tool, the Empirical Cross-Entropy (ECE) plot, previously used in forensic science [16,25] for two-class probabilistic classifiers.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have