Neural network based supervised rank aggregation

Abstract

Rank Aggregation problem is to find a combined ordering for objects, given a set of rankings obtained from different rankers. Rank aggregation is a technique that combines results of various rankings on the sets of entities (e.g. Documents or web pages of search result) to generate an overall ranking of the entities. In the context of the World Wide Web, Rank aggregation is frequently used in metasearching. In this paper, we discuss the development of a supervised rank aggregation system that is based on “neural networks”. A supervised rank aggregation system provides an aggregation of rankings of entities by learning rules for combining the different individual rankings of the entities on the basis of training data. In case of a metasearch system, the training data may be the user feedback based ranking of the search results. The main contribution of the paper is the formulation of the rank aggregation problem as a function approximation problem. As the multilayer perceptrons are considered as the universal approximator, we use a multilayer perceptron for the supervised rank aggregation. For experimental purpose, we apply this supervised rank aggregation technique to metasearching. We train our metasearch system with the user feedback based ranking of the search results from 7 public search engines for a set of 15 queries. We also evaluate the performance of our trained metasearch system using the feedback from three independent evaluators and find that our system gives a very good performance.