Title: "Evolutionary High-Dimensional Outlier Detection."
Abstract:
Outlier detection is not only a standard step in data pre-processing, but provides interesting insight into the data. There are many industries which utilize outlier (or anomaly) detection: credit card fraud, sports statistics, signal error identification, and health monitoring are a few notable examples. It is also often of interest to know the features of detected outliers, and whether there are particular variables that are unusually outlier-prone. Standard outlier detection methods utilize concepts of distance, which tend to fail under the "curse of dimensionality" and also lose the concept of spatial direction; sparsity can retain direction, but still breaks down in high dimensions, so dimension reduction is often required. In this talk, I will focus on an algorithm presented by Aggarwal and Yu (2004) which searches for sparsely populated regions on a lower-dimensional projection space using a computationally-efficient evolutionary approach.