Yi-Wei Liu - Colloquium Speaker
Everyone is welcome to join via Zoom:
3:00 – 3:30 pm CDT Meet & Greet
3:30 – 4:30 pm CDT Talk – “Change-point Detection for Modern Data”
Graph-based change-point detection is a nonparametric framework that utilizes the edge-count information on similarity graphs constructed among observations, which can be applied to high-dimensional/non-Euclidean data observations with analytical formulas to control the type I error. As we enter the era of big data, many challenges arise facing the demand of modern data analysis. For example, improving the efficiency of the algorithm when applied to large and complex datasets, controlling the false discovery rate for serial correlated data, and detecting systematic change-points in multiple sequences of observations. In this talk, I will present three variations of the graph-based methods that tackle these challenges. The first variation uses the directed approximate k-NN information to improve the efficiency of the algorithm. The second variation incorporates the circular block permutation scheme to achieve a more accurate type I error control for locally dependent data. In the third variation, we propose a new test statistic that can detect systematic change-points in multiple sequences. The new test not only has higher power but is also successful in detecting various types of changes.