Query difficulty prediction for contextual image retrieval

Abstract

This paper explores how to predict query difficulty for contextual image retrieval. We reformulate the problem as the task of predicting how difficult to represent a query as images. We propose to use machine learning algorithms to learn the query difficulty prediction models based on the characteristics of the query words as well as the query context. More specifically, we focus on noun word/phrase queries and propose four features based on several assumptions. We created an evaluation data set by hand and compare several machine learning algorithms on the prediction task. Our preliminary experimental results show the effectiveness of our proposed features and the stable performance using different classification models. © 2010 Springer-Verlag Berlin Heidelberg.