Building statistical models by visualization

Thomas P. Minka

Practical applications of statistical methods usually involve assumptions about the domain, such as independence, normality, linearity, and the choice of variables. I will describe a set of graphical tools which simplify the process of learning about a domain and checking model assumptions. Some of these tools have been around for years, but vastly underused by statistical learning researchers. Using real-world examples, I will illustrate how visualization can be used for model selection, identifying exceptional cases, and interpreting the results of learning algorithms.

Talk slides, pdf (221K)

Also see my page on Discriminative projections and the class materials on my main page.

My visualization software is available: (R 1.7.1) (R 1.9.1)
In R for Windows, select menu option

Packages -> Install Package from local zip file...
Then type

These packages only work with the version of R listed. Under Windows Vista, R has trouble unzipping packages. In this case, you need to manually unzip the package into the R/library directory.

Last modified: Wed Apr 26 12:39:31 GMT 2006