DMML Program 2003-04

The Support Vector

Machines Group

This cluster will investigate the problem of feature selections for SVM methods. Di Cook might have data for this topic. Maybe also think about missing values problems (in high p) for SVM? Case of mixed data. Interpretability of SVM methods. From a very informal brainstorming session, various questions arised, among which the following: (1) How does Kernel help feature selection. What is the connection/mechanism between Kernel and feature selection. How does feature selection affect SVM classification. One of possible goal is to develop a model for simultaneous feature selection and classification. (2) Kernel PCA is a generalization of PCA in the SV favor. How to choose the best Kernel for Kernel PCA remains a open question. We will study the intrinsic relations between Kernel PCA and SVM to connect feature selection and SVM. (3) Feature selection and SVM on short fat data. We can investigate the missing value problems and data with mixed types in this setting. With models being developped, Di's data sets can be used for experimental studies.

• Next meeting: Wed, Sep 24, at 3:00pm
• There have been many references and software packages contributed. Please these out and send your feedback.
• There will be a series of presentations from October 1st onwards. If you haven't yet volunteered to present, please start thinking of insights you can share in the form of a presentation to the group.

Members of the team