872488f88d1b2db54d55bc8bba2fad1b-Reviews.html

The authors propose a model of similarity in which a binary similarity score between two vectors (u,v) is modeled as a noisy-or over a set of (learned) Mahalanobis distances. The motivation being that abstract "similarity" is often non-transitive, and is not well-modeled by a single, fixed distance metric over a vector space. Generally speaking, the paper is clearly written and the method makes intuitive sense. The algorithm seems to work in practice, although the experimental results could be expanded a bit to better illustrate the method. Overall, I enjoyed reading this paper.