A0310
Title: Explainable outlier detection based on Shapley values
Authors: Marcus Mayrhofer - TU Wien (Austria) [presenting]
Peter Filzmoser - TU Wien (Austria)
Abstract: Shapley values are a popular method in Explainable AI used to quantify the contribution of each feature to the model outcome for each observation. Recently, their practicality has been demonstrated in the context of multivariate outlier detection to explain why an observation deviates from the majority of the data. The Shapley values are used to decompose the squared Mahalanobis distance into outlyingness scores for each variable, which can be interpreted as average marginal contributions to the outlyingness of an observation. The additivity property of the Shapley value ensures that the sum of those contributions is identical to the squared Mahalanobis distance. Although the computational complexity of the Shapley value is a drawback in general, it can be reformulated and simplified in our case. This reformulation enables quick and efficient computation even in higher dimensions. Multivariate anomaly explanation can be extended to the setting of matrix-variate observations and provides a different perspective on cellwise outlyingness, which aims to detect outlying cells within a data matrix.
