Visual Analytics - Cluster Analysis (1)

Lecture four, consisting of two parts as well, and is also dedicated to clustering. Here we discuss the validation of clustering results and the visualization of (global) clustering results. Validation means to assess the quality of a clustering. Did we summarize instances in one cluster that better should be separated? Or are there instances that should be in the same cluster, but are not? The purity of a cluster, or the silhouette coefficient are measures for the cluster quality that do not require a ground truth. It is tempting to just compute one such measure and to trust it. We give examples when such measures are misleading. As a consequence, several measures should be computed, and the clustering results should also be visualized. Isolines, dendograms, glyphs, heatmaps, as well as 2D and 3D scatterplots are among the used visualization techniques. It is often necessary to combine subjective and qualitative assessment with measured quality aspects. Chapters: 00:00 - Outline and Introduction 09:30 - Silhouette Coefficient 17:02 - Centroid-Based Measure 25:14 - Grid-Based Method 31:57 - Visualization of Clustering Results