Events — Colloquia & Seminars
Learning in the presence of class imbalance
Speaker: Nitesh Chawla
Date: Tuesday, May 6, 2008
Talk: 12:00 PM, 366 WVH
Abstract
Models for knowledge discovery in the real world face the pervasive and compelling problem of irregularities in data distribution. Decisions that are optimal in expected utility can be vulnerable to catastrophic failure, and value functions that reflect the discontinuities of the real world pragmatics can quickly become intractable. Surprises can happen in uncertain environments. The class distributions may not be the same, with the class of interest being rare. The training and testing distributions can differ. The costs of making mistakes or benefits from making correct predictions may also not be constant and can evolve due to operational reasons. In this talk, I will present our work on mining in the presence of unbalanced data, as distributions shift.
Brief Biography
none provided