A relevant score normalization method using shannon's information measure

Authors:
Yu Suzuki;Kenji Hatano;Masatoshi Yoshikawa;Shunsuke Uemura;Kyoji Kawagoe
Affiliations:
Ritsumeikan University, Shiga, Japan;Nara Institute of Science and Technology, Nara, Japan;Nagoya University, Nagoya, Aichi, Japan;Nara Institute of Science and Technology, Nara, Japan;Ritsumeikan University, Shiga, Japan
Venue:
ICADL'05 Proceedings of the 8th international conference on Asian Digital Libraries: implementing strategies and sharing experiences
Year:
2005

Citing 3
Cited 1

Relevance score normalization for metasearch

Proceedings of the tenth international conference on Information and knowledge management
Condorcet fusion for improved retrieval

Proceedings of the eleventh international conference on Information and knowledge management
Features in Content-based Image Retrieval Systems: a Survey

State-of-the-Art in Content-Based Image and Video Retrieval [Dagstuhl Seminar, 5-10 December 1999]

Using knowledge integration techniques for user profile adaptation method in document retrieval systems

Transactions on computational collective intelligence V

Quantified Score

Hi-index	0.00

Visualization

Abstract

Given the ranked lists of images with relevance scores returned by multiple image retrieval subsystems in response to a given query, the problem of combined retrieval system is how to combine these lists equivalently. In this paper, we propose a novel relevance score normalization method based on Shannon's information measure. Generally, the number of relevant images is exceedingly smaller than that of the entire retrieval targets. Therefore, we suppose that if the subsystems can clearly identify which retrieval targets are relevant, the subsystems should calculate high relevance scores to a few retrieval targets. In short, we can calculate the sureness of the IR subsystem using the distribution of the relevance scores. Then, we calculate the sureness of the IR subsystems using Shannon's information measure, and calculate the normalized relevance scores using the sureness of the IR subsystems and the raw relevant scores. In our experiment, our normalization method outperformed the others.