A Decomposition of Total Variation Depth for Understanding Functional Outliers

There has been extensive work on data depth-based methods for robust multivariate data analysis. Recent developments have moved to infinite-dimensional objects, such as functional data. In this work, we propose a notion of depth, the total variation depth, for functional data, which has many desirable features and is well suited for outlier detection. The proposed depth is in the form of an integral of a univariate depth function. We show that the novel formation of the total variation depth leads to useful decomposition associated with shape and magnitude outlyingness of functional data. Compared to magnitude outliers, shape outliers are often masked among the rest of samples and more difficult to identify. We then further develop an effective procedure and visualization tools for detecting both types of outliers, while naturally accounting for the correlation in functional data. The outlier detection performance is investigated through simulations under various outlier models. Finally, the proposed methodology is demonstrated using real datasets of curves, images, and video frames.