Logistic Regression Revisited: Belief Function Analysis

We show that the weighted sum and softmax operations performed in logistic regression classifiers can be interpreted in terms of evidence aggregation using Dempster's rule of combination. From that perspective, the output probabilities from such classifiers can be seen as normalized plausibilities, for some mass functions that can be laid bare. This finding suggests that the theory of belief functions is a more general framework for classifier construction than is usually considered.