Export Reviews, Discussions, Author Feedback and Meta-Reviews

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This work address the problem of learning a ranking prediction function that optimizes (N)DCG. The authors propose a surrogate loss based on a non-convex upper bound of the DCG, inspired from robust classification losses. The difference with other existing non-convex upper-bound resides in the fact that the authors introduce the non-convexity at the context level (on a whole query) and not at the pair of items level (see [8]). Then, the authors propose two applications of their algorithm with experimental studies: one is learning a prediction model for a search engine problem, the other to learn a representation for collaborative filtering.