Enhancing a Ranking System Practical Applications to Information Retrieval Systems and BCS Rankings

Abstract: The rankings of the members fora given domain, if provided, can be thought to be the efficient description of the domain and help us make decisions accordingly. For example, each web search engine like Google retrieves query- relevant documents and shows us highly ranked ones in ascending order of their ranks. However, if we have two or more rankers or ranking systems with their own expertise for one domain, we may be puzzled what to choose or how to fix their differences. In this paper, the issue of how to combine two experts’ rankings is treated. We suggest some methods of combining two different rankings. For practical applications, two domains were selected to test and validate our method. First, two different rankings generated by changing some portion of our information retrieval system were selected and our experiments show that the resulting average rank of top ten relevant documents, for example, was considerably improved. The second domain we tried may be thought to be almost improbable. The question is if American college BCS football ranking, officially published from the middle of the football season, can be modified to better predict the four week later BCS ranking itself. Our experiments show that some computer-based ranker help enhance the predictability of the future BCS ranking.

Keywords: ranking, machine learning, information retrieval, search engines, TF-IDF, NCAA football, computer rankings, BCS ranking