Visual Analytics - Subspace Clustering (1)

In this lecture, we continue with discussing clustering variants. Instead of global clusterings that suffer from the curse of dimensionality, i.e., in high-dimensional data, basically all elements are far from each other, subspace clustering identifies clusterable subspaces and applies clustering in them. A clusterable subspace is a subspace with a strongly varying density of the data, e.g., the distances to nearest neighbors vary considerably. Again, different methods are employed and are selected based on the properties of the data. As examples, we discuss grid-based methods and explain the „Ranking Interesting Subspaces“ algorithm (Kailing, 2003) in more detail, including a discussion of its parameters. SUBCLUE, Proclus and CLIQUE are further essential algorithms. The visual exploration of subspace clustering results is challenging: Subspace clusters may overlap in the instances and in the dimensions – the amount of this overlap needs to be visually represented. Overview visualizations should also indicate how similar subspace clusters are to each other. We explain these and other requirements for the visual exploration before we discuss specific solutions. Detail views complement them by showing distributions of values for a selected cluster. We discuss several visualization techniques, such as scatterplots, parallel coordinates, graph-based views, and multiple linked views. Chapters: 00:00 - Outline and Introduction 08:46 - Subspace Search 24:50 - RIS and Surfing 39:15 - Subspace Clustering 44:16 - Clique, SubClue, Proclus 52:10 - Parameters and Visualization