Credible classification for environmental problems

Abstract

Classifiers that aim at doing credible predictions should rely on carefully elicited prior knowledge. Often this is not available so they should start learning from data in condition of near-ignorance. This paper shows empirically, on an agricultural data set, that established methods of classification do not always adhere to this principle. Traditional ways to represent prior ignorance are shown to have an overwhelming weight compared to the information in the data, producing overconfident predictions. This point is crucial for problems, such as environmental ones, where prior knowledge is often scarce and even the data may not be known precisely. Credal classification, and in particular the naive credal classifier, is proposed as more faithful ways to cope with the ignorance problem. With credal classification, conditions of ignorance may limit the power of the inferences, not the credibility of the predictions.