A measure of surprise for incongruence detection

The ability to detect unexpected events improves dramatically when more than one expert is involved in decision making. Incongruence between two or more experts is indicative of something unusual and its measuring has applications in domains such as anomaly detection and multimodal decision systems. In this paper, we propose a new classifier incongruence measure, which overcomes the critical shortcomings of those existing in the literature. An experimental study has been carried out showing the advantageous properties of the proposed measure including its relatively low sensitivity to estimation noise, under the assumption of constrained Gaussian distribution. For different noise-free measure values corrupted with different levels of noise, we show that it is possible to determine classifier incongruence thresholds at given levels of statistical significance.