2000
@inproceedings{MMM2000d,
vgclass = {refpap},
vgproject = {viper,cbir},
author = {Henning M\"{u}ller and Wolfgang M\"{u}ller and David McG.\
Squire and St\'ephane Marchand-Maillet and Thierry Pun},
title = {Learning Feature Weights from User Behavior in Content-Based
Image Retrieval},
booktitle = {MDM/KDD2000 Workshop on Multimedia Data Mining in
conjunction with the Sixth ACM SIGKDD International Conference on
Knowledge Discovery \& Data Mining},
address = {Boston, USA},
month = {August~20},
year = {2000},
url = {/publications/postscript/2000/MuellerHMuellerWSquireMarchandPun_mdm2000.pdf},
url1 = {/publications/postscript/2000/MuellerHMuellerWSquireMarchandPun_mdm2000.ps.gz},
abstract = {This article describes an algorithm for obtaining
knowledge about the importance of features from analyzing user log
files of a content-based image retrieval system (CBIRS). The user log
files from the usage of the \emph{Viper} web demonstration system are
analyzed over a period of four months. Within this period about 3500
accesses to the system were made with almost 800 multiple image
queries. All the actions of the users were logged in a file.
The analysis only includes multiple image queries of the system with
positive and/or negative input images, because only multiple image
queries contain enough information for the method described. Features
frequently present in images marked together positively in the same
query step get a higher weighting, whereas features present in one
image marked positively and another image marked negatively in the same
step get a lower weighting. The \emph{Viper} system offers a very
large number of simple features. This allows the creation of flexible
feature weightings with high values for important and low values for
less important features. These weightings for features can of course
differ between collections and as well between users. The results are
evaluated with an experiment using the relevance judgments of real
users on a database containing 2500 images. The results of the system
with learned weights are compared to the system without the learned
feature weights.},
}