Monday, April 20, 2009
Ke Dang
Advisor: James French; Worthy Martin
Attending Faculty:
OLSSON 236D, 10:00:00
A Master's Project Presentation
Learning to Rank from a "Base Model"
ABSTRACT
In my project I propose an algorithm that attempts to make better use of the many document models
and features that have been created over the last several years in information retrieval research.
My algorithm first selects a "base model" from the complete set of available features based on
performance on a training set. From that performance, the algorithm assigns a weight to the
"base model". The algorithm then assigns weights to remaining document models and features
according to their relative performance to the "base model". The ranking of the documents in the
"response set" for a given query is then established by a linear combination (via the established weights)
of the features measured on those documents. The algorithm has the advantages of having a rather straight-forward implementation, the training process is tractable and the incorporation of new document features is simple. I will report on experiments over the OHSUMED document set that indicate that the algorithm is
comparable to the baseline Adarank algorithm.