Regression on LETOR 

 

Introduction

Learning parameters

Notes

 

Introduction

The aim of this entry is to provide a competitive baseline in the pointwise approach category. Compared to plain regression, this entry includes two rather standard enhancements:
  • A regularizer; in other words, we perform ridge regression.
  • A reweighting in such way that the relevant and non-relevant documents have the same total weight. In the context of SVM classifier, this trick has for instance been used by [Morik et al. '99] (see equation (4)).

Learning Parameters

The details of the implementation can be found by looking at the Matlab source code.

In particular, the ridge is chosen to be a constant times the mean value of the diagonal elements of the covariance matrix; and this constant is selected on the validation set in the set

{0.0001, 0.001, 0.01, 0.1, 1, 10}.

Dataset

Regularization constant (from Fold1 to Fold5)

OHSUMED

1, 1, 0.1, 0.1, 0.1

TD2003

0.01, 0.001, 0.1, 0.1, 0.001

TD2004

0.0001, 0.01, 0.0001, 0.01, 0.01

HP2003

0.1, 0.1, 0.1, 0.1, 1

HP2004

1, 10, 0.01, 0.1, 1

NP2003

1, 0.01, 0.1, 0.1, 0.01

NP2004

1, 10, 10, 1, 10

Notes

This document was written by Olivier Chapelle, and the experiments were conducted by Olivier Chapelle. If any problem, please contact letor@microsoft.com or chap@yahoo-inc.com