Computing That Serves

Which Supervised Learning Method Works Best for What? An Empirical Comparison of Learning Methods and Metrics++


Thursday, March 1, 2007 - 11:00am


Rich Caruana, Department of Computer Science, Cornell University

Decision trees are intelligible, but do they perform well enough that you should use them?  Have SVMs replaced neural nets, or are neural nets still best for regression, and SVMs best for classification? Boosting maximizes margins similar to SVMs, but can boosting compete with SVMs?  And if it does compete, is it better to boost weak models, as theory might suggest, or to boost stronger models?  Bagging is simpler than boosting -- how well does bagging stack up against boosting?  Breiman said Random Forests are better than bagging and as good as boosting.  Was he right?  And what about old friends like logistic regression, KNN, and naive bayes?  Should they be relegated to the history books, or do they still fill important niches?
In this talk we compare the performance of these supervised learning methods on a number of performance criteria: Accuracy, F-score, Lift, Precision/Recall Break-Even Point, Area under the ROC, Average Precision, Squared Error, Cross-Entropy, and Probability Calibration. The results show that no one learning method does it all, but some methods can be "repaired" so that they do very well across all performance metrics.  In particular, we show how to obtain the best probabilities from max margin methods such as SVMs and boosting via Platt's Method and isotonic regression.  We then describe a new ensemble method that combines select models from these ten learning methods to yield even better performance.  Although these ensembles perform extremely well, they are too complex for many applications. We'll describe a model compression method we are developing to fix that.  Finally, if time permits, we'll discuss how the performance metrics relate to each other, and which of them you probably should (and shouldn't) use.


Rich Caruana is an Assistant Professor of Computer Science at Cornell University.  He got his Ph.D. at CMU in 1997 where he worked with Tom Mitchell and Herb Simon.  Before joining the faculty at Cornell in 2001 Rich was a researcher at Just Research (JPRC) and visiting faculty in the Medical School at UCLA and at CMU's Center for Learning and Discovery (CALD).  Rich's research is in machine learning and data mining, and applications of these to medical decision making, bioinformatics, and weather forecasting.  He is best known for his work in inductive transfer, semi- supervised learning, and optimizing learning for different performance criteria.  Rich likes to mix algorithm development with applications work to insure that the methods he developes really work in practice.