Visual Analytics - Visual Analysis of Biclusters (1)

This video lecture explains and motivates the use of bi-clustering. This lecture concludes the clustering part of the course. Bi-clustering (or co-clustering) is a special instance of subspace clustering. It is based on the mathematical concept of bi-partite graphs. The resulting clusters are rectangular in shape, they are determined in 2D data, representing, e.g., students that take the same courses. Thus, bi-clustering is applied to categorical data as they occur, for example, in crime analysis. More recently, it was extended to numerical data, where instead of equality we can check whether the difference between two values is below a certain threshold. This kind of bi-clustering is primarily used for gene-expression data, i.e., in bioinformatics. To reduce the dependency of a threshold, fuzzy bi-clustering was introduced (FABIA algorithm). As always, we are not only interested in analytical techniques: we want to integrate them in a visual analytics process, i.e., we have to consider pre-processing, e.g., noise removal and after clustering we have to visualize the results, and we should support the interactive exploration and sense-making. Bi-clustering results may overlap – datasets may belong to different bi-clusters. This overlap needs to be faithfully conveyed. The results of bi-clustering may be overwhelming. Thus, we discuss interestingness measures to reduce the results, e.g., to avoid very small clusters. Chapters: 00:00 - Outline and Introduction 13:22 - Definition and Algorithms 22:26 - Preprocessing and Postprocessing 25:32 - Tasks 28:08 - Parameters 34:47 - Interaction and Visual Exploration