NeurIPS 2020

Learning from Label Proportions: A Mutual Contamination Framework

Review 1

Summary and Contributions: The authors study the problem of learning from label proportions by using mutual contamination models. To this end, they propose a new unbiased loss and are able to derive generalization bounds under non-iid samples. A weighted empirical risk then arises naturally that is consistent for LLP. Experiments are simulated where there is an important improvement in the scenario of larger bag sizes.

Strengths: 1. The theoretical contributions are sound. I checked some of the proofs in more or less detail and could not spot any errors. 2. The idea of mixing MCM with LLP is a novel contribution and authors derived generalization bounds and proposed a consistent learning procedure. 3. The content is mostly theoretical with some experiments that support the theory. Thus, the work is relevant to the learning theory community and can also be of interest to practitioners.

Weaknesses: * There are no comments on the computational complexity of their method and how it compares to previous work. * There are no comments on the tightness or optimality of the generalization bound.

Correctness: Yes, theoretical and empirical results are sound.

Clarity: Yes, the paper is well written. I enjoyed reading it.

Relation to Prior Work: Authors do a good job in highlighting the difference between each of their contributions and existing work.

Reproducibility: Yes

Additional Feedback: 1. Given that their method uses bag-pairing, k-merging schemes. I would like to see some discussion on the computational complexity and how it compares to previous methods. For instance, can this method be scaled? (My guess is that it could be difficult given the use of ILP) 2. Optimality of bounds can be another paper on its own, but do the authors have some comments on it? 3. I might be missing something here, but in Line 292 I read that LLPs are uniform on [0, 0.5] and on [0.5, 1]? Does this mean that in the experiments we do not have, say, a label proportion of 0.2 and another label proportion of 0.8? If this is the case, why? 4. Would be worth looking at the precision-recall AUC in those cases? I am curious what would be the performances. === UPDATE After rebuttal, I am keeping my score unchanged. It's not clear if the results would hold when using approximate solvers for ILP in the case of unequal sized bags. Thus, I still have a concern on the practical usability of the method *with* the given guarantees and cannot vote for a greater score.

Review 2

Summary and Contributions: The paper encapsulates learning from label proportions in a more general framework, called mutual contamination framework. It is shown that if one knows the contamination rates (akin to the proportions in the LLP framework), then the models can be learned to minize the EBR loss over instances. Concentration inequalities in the fashion of Rademacher complexity are then given to show that efficient training can take place in practice, with the bounds suggesting ways to make optimal choices in the method parameters. Experiments show that the method performs quite well, in particular with large-sized bags.

Strengths: * An encompassing framework for LLP, allowing to derive consistency results under quite general conditions. * Corresponding learning algorithms that show a quite good performances.

Weaknesses: * Not much, the paper is well-written, with a first part accessible to a wide audience, and a second part focusing on technical results of the paper. As said by the authors, one could think of a multi-class setting, yet the binary case is an essential and important first step towards it.

Correctness: * I am not an expert in Rademacher complexity, but the results appeared correct to me.

Clarity: * Although the second part gets quite technical, the paper is overall well-written, with clear explanations about the interest and limitations of the results.

Relation to Prior Work: * Yes

Reproducibility: Yes

Additional Feedback: * Unfortunately for the authors, this is a bit far from my expertise to give meaningful comments (I could not check this before, due to the COVID situation).

Review 3

Summary and Contributions: The paper deals with a weakly supervised learning problem referred to as Learning from Label Proportions (LLP), akin to standard multiclass classification except that data are grouped into datasets (‘bags’) and only the class proportions are at disposal for learning. The approach developed in the article is that of Mutual Contamination Models (MCMs). It is described in section 2 in the binary case: data within a bag being generated from a certain mixture of the two class distributions, bags being independent from each other. Two situations are considered, depending on whether the data within a bag are independent or not. The goal pursued is to minimise the balanced error rate (BER), i.e. the error rate when the parameter mixture is equal ti 1/2. Generalisation bounds are obtained in terms of BER and involve specific complexity measures described in Definition 3. In 3.1, it is explained that, when the true proportions are known that LLP boils down to applying results for MCMs by pairing bags, whereas the situation where only empirical proportions are available is considered in 3.2: a maximal deviation bound is stated in Theorem 6. Experimental results are displayed in section 4.

Strengths: The LLP framework is well motivated, the results seem to be novel and the parallel drawn with MCMs may be original.

Weaknesses: Regarding the form, the paper could be improved by being less dense. Although none of the arguments is complex, the paper is very hard to follow. A very general framework is introduced (cf dependence structure), whereas restrictive assumptions are next made (cf conditional independence): it is very difficult to know which hypotheses cannot be avoided. In addition, the form of the generalisation bounds is weird, because not formulated in terms of excess of risk, as if nothing could be said on the optimal learning strategy in the LLP framework. And the choice of BER seems arbitrary as well, insofar as a ROC analysis of the decision functions output by the algorithm proposed is carried out in the experimental section, The BER being just the AUC of the binary classifier, corresponding to a single point of the ROC curve of the decision function (where the tangent has slope equal to one), there is no reason for the whole ROC curve to be accurate everywhere.

Correctness: The paper seems correct to me but I am skeptical of the relevance of the BER analysis carried out here.

Clarity: The paper is understandable but could be improved (style and general organisation).

Relation to Prior Work: The state of the art is well described.

Reproducibility: Yes

Additional Feedback:

Review 4

Summary and Contributions: This paper provides a novel solver for LLP problem based on unbiased surrogate loss for balanced error rate (BER). The main contribution is on theorical side. I have checked the proofs in a general way, and I don’t find any obvious mistake. I think it will be better if the authors can summarize their contributions in the introduction part clearly, it will help the readers to catch the main results of this paper.

Strengths: although the authors claim that “the theoretical underpinnings of LLP have been slow to develop”, and they offer “a general-purpose, theoretically grounded empirical objective for training LLP classifiers”, it seems that they only extend the former mutual contamination model to the multiple models scenario, to seamlessly apply the results to solve LLP problems. I cannot find any novel insights on LLP specifically. As a result, I think this paper is to offer a new algorithm for LLP with better theorical grounds, compared with the existing LLP solvers.

Weaknesses: 1. The work is strongly based on the results for mutual contamination models, which is specially designed and discussed for binary classification problems. As a result, the benchmark approaches involved in the experimental section are confined in two early proposed methods for binary problems. Many up-to-date models that focus on multi-class LLP problems are not touched by this work. In other words, there is a gap between this work and multi-class LLP problem. In particular, thanks to deep neural networks, the performance on LLP has been greatly improved by recent work, such as the work in [1], [11], [18], [31], and [34], on much complicated datasets, e.g., image data. This is an obvious limitation of this work. The authors should notice this shortcoming and continue to provide an explanation or fix this problem in their future work. 2. Even for the binary case, the experiments are insufficient in this paper, and the results is not significantly better than that of the two algorithms, given they are out-of-date in solving LLP. I suggest adding more experimental results. For example, more experimental results on hyper-parameter analysis. In addition, for the K-merging schemes, I do not quite understand how it work and how the number K is fixed in the experiments. Please offer more insights or analysis on this issue for improvement.

Correctness: Yes

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: