NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2956
Title:A Refined Margin Distribution Analysis for Forest Representation Learning

Reviewer 1


		
Originality: To the best of my knowledge, the work presented in this paper is quite novel. Clarity: The paper is well written and well structured. Quality: The paper advances the understanding of CasDF with strong well proved theoretical results. Moreover, the derived algorithm makes sense, and the experimental analysis is strong enough. Significance: The theoretical result of the paper is in my opinion important and useful to the field. The experiments show improvement over the baselines in all setting. One missing detail in the results is to report average performance with standard deviation to show the significance of this improvement, especially that some results are close to each other.

Reviewer 2


		
[response to authors] I’ve read all reviewers’ comments and the authors’ response letter. I think this work has theoretical novelty in understanding CASDF and has experimental support. It will be valuable to the field for future studies. Comments are addressed. Overall, I will keep my score. -------------------------- This paper provides a new perspective to understand and explain the cascade deep forest, and proposes a margin distribution reweighting approach to minimize the gap between the generalization error and empirical margin loss, which produces much better model performance. The method is supported by mathematical theories and empirical experiments. The theories and proofs are clear and solid. Experiment settings and results are clear.

Reviewer 3


		
Originality: The primary contribution is a tighter bound. From skimming, this comes from using a bernstein style bound so when the variance is small, there can be a tighter rate. Quality: The paper is reasonably well written and the claims appear correct. I did not check the proofs. Clarity: The paper is reasonably well-organized. It is still difficult to follow because it requires the reader to be familiar with Zhou and Feng though it is concisely summarized. Significance: Unclear. The primary contribution seems to be an improved boosting style algorithm that removes a sqrt. The experiments are completely unconvincing.