Neighbor-aware Geodesic Transportation for Neighborhood Refinery

Jifei Luo, Xiao Cui, Hantao Yao, Changsheng Xu

Keywords: image retrieval, reranking, deep clustering, self-supervised learning

Abstract:

Neighborhood refinery aims to enhance the neighbor relationships by refining the original distance matrix to ensure pairwise consistency.
Traditional context-based methods, which encode instances alongside their local neighbors in a contextual affinity space, are limited in capturing global relationships and are vulnerable to the negative impacts of outliers in the neighborhood. To overcome these limitations, we propose a novel Neighbor-aware Geodesic Transportation (NGT) for the neighborhood refinery. NGT first constructs a global-aware distribution for each instance, capturing the intrinsic manifold relationships among all instances. This is followed by an optimization transportation process that utilizes the global-aware distribution within the underlying manifold, incorporating global geometric spatial information to generate a refined distance. NGT first involves Manifold-aware Neighbor Encoding (MNE) to project each instance into a global-aware distribution by constraining pairwise similarity with the corresponding affinity graph to capture global relationships. Subsequently, a Regularized Barycenter Refinery (RBR) module is proposed to integrate local neighbors into a barycenter, employing a Wasserstein term to reduce the influence of outliers. Lastly, Geodesic Transportation (GT) leverages geometric and global context information to transport the barycenter distribution along the geodesic paths within the affinity graph. Extensive evaluations on several tasks, such as re-ranking and deep clustering, demonstrate the superiority of our proposed NGT.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7466