Varying parameters of a clustering algorithm or applying several algorithms to the same data produces different partitions of the data. Comparing these partitions is essential to understanding the underlying structure in data. However, such comparisons, if done, are often done in an ad-hoc way. We present Coral: an application for exploration and quantitative and qualitative analysis of clustering ensembles. Coral makes all-to-all clustering comparison easy, supports exploration of individual clusterings, allows tracking cluster dynamics across partitions, and supports identification of core and peripheral items in the dataset.
Darya Filippova. Exploring multiple network partitionings with Coral. Tech report, 2009. University of Maryland, College Park.
Participants: Darya Filippova, Aashish Gadani, Megan Riordan, Carl Kingsford
Coming
(more)
Explore and Visualize Crashes (EVC) online tool serves as a crash reference for traffic engineers in the state of Maryland. Police crash records contain a vast amount of information describing almost every aspect of the collision, however, mining that data requires extensive knowledge of querying languages and familiarity with the data itself. EVC is the first application that allows engineers to interactively query for specific crash records and analyze them through a variety of custom visual displays.
Participants: Darya Filippova, Sarah Rheel, Michael Pack
tech report
Analyzing multivariate data sets is not an easy task: the user has to analyze individual variables as well as the relationships between them. Lower-level projection techniques may assist the user in finding interesting combinations; however, few tools support systematic exploration of those combinations. One way to deal with the "curse of dimensionality" is to rank all such relationships according to some measure of interestingness. At the same time, clustering and ranking algorithms are well researched for continuous data, while categorical data analysis has not received equal attention. This paper explores the ways to analyze categorical data sets and visualize and rank the relationships between categorical variables. CateRank uses histograms (bar charts) to visualize one-dimensional variable distribution and reorderable matrix to visualize the relationship between two categorical variables. The tool proposes several metrics based on the matrix properties that describe the nature of the relationship between the two categorical variables and allow comparing relationships within the data set.
Participants: Darya Filippova, Ben Shneiderman
HCIL webpage