Seminar Details

Seminar Details


Monday

May 24

3:30 pm

Boosting Algorithms and Predictive Modeling

Yoav Freund

Seminar

Columbia University

Modern machine Learning algorithms such as support vector machines and boosting are rapidly becoming standard tools in the analysis of large and complex data-sets. Examples include the analysis of transaction records in business data mining, the analysis of gene regulatory networks in computational biology, event classification in experimental high-energy physics and object detection in computer vision.

The novelty of these methods is both statistical and computational. From the statistical point of view, these new methods perform surprisingly well in situations where the dimension of the data is far larger than the number of available data-points. This phenomenon can be explained in terms of the prediction "margins" and their relationship to prediction confidence.

From the computational point of view, the main novelty of boosting is the discovery of an efficient way in which many weakly correlated features can be combined into a single accurate predictor.

There is an underlying paradigm shift in statistical analysis. Traditionally, the main goal of the statistician is to accurately estimate the parameters of a stochastic model of the system under investigation and thereby add directly to scientific knowledge. However, in an increasing number of applications, the goal is to construct a prediction function, implemented as a computer algorithm, which can predict some important aspect of the system. An intuitive interpretation of the parameters of this prediction function is of secondary importance. This brings machine learning methodology into direct conflict with the prevailing scientific research paradigm.

In this talk I will describe boosting algorithms, their theoretical analysis, some of their more interesting applications, and some of the directions of improvement. I will then describe some recent work on using boosting as a way of improving the efficiency of data collection. I'll conclude with some observations regarding the increasing need to interpret prediction functions and incorporate them into humanly understandable scientific knowledge.