Google 2005

Logistic regression for fast, accurate, and parameter free data mining

(alternately, Logistic Regression: Not Dead Yet)

I recently gave a talk at Google, Inc. The slides below correct a couple minor omissions from the presentation, and a few typos. Use the navigation bar at the bottom to see my other papers or download datasets. NOTE: corrected again (minor) on 30 Nov 2005.

Corrected slides: google2005.pdf

Abstract

Logistic regression (LR) is a well-known and well-understood regression method. It is commonly used in medical research to produce explanatory models, but can be used as a fast and accurate binary classifier in data mining applications. Support vector machines (SVM) are often considered state-of-the-art for labeled classification tasks, but LR can be just as accurate and significantly faster than popular SVM implementations. In this talk, we will briefly describe our simple LR implementation, discuss usual and unusual applications of LR, and speculate about how large of classification problems LR can solve on modern computing equipment.