- This event has passed.
Advances in Regularization: Bridge Regression and Coordinate Descent Algorithms
June 4, 2012 @ 6:30 pm - 8:00 pm
ACM Data Mining SIG
Title: Advances in Regularization: Bridge Regression and Coordinate Descent Algorithms
Speaker: Giovanni Seni
Please note the date change to June 4
A widely held principle in Statistical model inference is that accuracy and simplicity are both desirable. But there is a tradeoff between the two: a flexible (more complex) model is often needed to achieve higher accuracy, but it is more susceptible to overfitting and less likely to generalize well. Regularization techniques “damp down” the flexibility of a model fitting procedure by augmenting the error function with a term that penalizes model complexity. Minimizing the augmented error criterion requires a certain increase in accuracy to “pay” for the increase in model complexity (e.g., adding another term to the model). This talk offers a concise introduction to this topic and a review of recent developments leading to very fast algorithms for parameter estimation with various types of penalties. It concludes with an example in R, showing an application of the techniques to a document classification task with 1-Million predictors.
Giovanni Seni is currently a Senior Data Scientist with Intuit. As an active data mining practitioner in Silicon Valley, he has over 15 years R&D experience in statistical pattern recognition, data mining, and human-computer interaction applications. He has been a member of the technical staff at large technology companies, and a contributor at smaller organizations. He holds five US patents and has published over twenty conference and journal articles. His book with John Elder, “Ensemble Methods in Data Mining – Improving accuracy through combining predictions”, was published in February 2010 by Morgan & Claypool. Giovanni is also an adjunct faculty at the Computer Engineering Department of Santa Clara University, where he teaches an Introduction to Pattern Recognition and Data Mining class.