CoSy Zoom seminar

Register here

Abstract:

Time series (TS) are a frequent data format occurring in many real life applications. In order to enable posterior analyses of this broad range of applications, unsupervised pattern mining approaches for TS become relevant. In this talk, TS segmentation for TS whose domains display distinctive and repeated patterns is addressed. Previous approaches utilizing the matrix profile (1NN graph) are briefly reviewed. By considering how arcs, a curve pointing from any index to its nearest neighbor, can be appropriately weighted to then be summed into an arc curve containing information about domain boundaries, some of the limitations of previous approaches are addressed. These findings are combined into a sliding window approach which is then demonstrated on data sets from different sensors - thus achieving a sensor/domain agnostic unsupervised TS segmentation.