Classifying Images with Image and Text Search Clickthrough Data

Abstract

Clickthrough data from search logs has been postulated as a form of relevance feedback, which can potentially be used for content classification. However there are doubts about the reliability of clickthrough data for this or other purposes. The experiment described in this paper gives further insights into the accuracy of clickthrough data as content judgement indicators for both HTML pages and images. Transitive clickthrough data based classification of images contained in HTML pages has been found to be inferior to direct classification of images via image search clickthrough data. This experiment aimed to determine to what extent this is due to the inferior accuracy of clickthrough-based classification accuracy in HTML. The better classifications resulting from clickthroughs on image searches is confirmed.