%0 Conference Proceedings %F liuw:RAP %A Liu, Wei %A Jiang, Wei %A Chang, Shih-Fu %T Relevance Aggregation Projections for Image Retrieval %B ACM International Conference on Image and Video Retrieval %C Niagara Falls, Canada %X To narrow the semantic gap in content-based image retrieval (CBIR), relevance feedback is utilized to explore knowledge about the user¡¯s intention in finding a target image or a image category. Users provide feedback by marking images returned in response to a query image as relevant or irrelevant. Existing research explores such feedback to refine querying process, select features, or learn a image classifier. However, the vast amount of unlabeled images is ignored and often substantially limited examples are engaged into learning. In this paper, we address the two issues and propose a novel effective method called Relevance Aggregation Projections (RAP) for learning potent subspace projections in a semi-supervised way. Given relevances and irrelevances specified in the feedback, RAP produces a subspace within which the relevant examples are aggregated into a single point and the irrelevant examples are simultaneously separated by a large margin. Regarding the query plus its feedback samples as labeled data and the remainder as unlabeled data, RAP falls in a special paradigm of imbalanced semi-supervised learning. Through coupling the idea of relevance aggregation with semi-supervised learning, we formulate a constrained quadratic optimization problem to learn the subspace projections which entail semantic mining and therefore make the underlying CBIR system respond to the user¡¯s interest accurately and promptly. Experiments conducted over a large generic image database show that our subspace approach outperforms existing subspace methods for CBIR even with few iterations of user feedback %U http://www.ee.columbia.edu/dvmm/publications/08/rap_civr08.pdf %8 July %D 2008