NeurIPS 2020

### Review 1

Summary and Contributions: This paper introduces a model for statistics for settings where standard assumptions like "data is i.i.d. from a distribution in a known class" may be (badly) violated. The authors give an algorithm for estimation in this model and prove that it incurs error comparable to the optimal algorithm in a natural but restricted class of estimators. The authors provide convincing numerical experiments demonstrating that their estimation algorithm outperforms standard strategies.

Strengths: Let me first summarize the paper somewhat more thoroughly. The model introduced by the authors is as follows. There are $n$ data points $\{x_1,\ldots,x_n\} \in [-1,-1]$; these may be arbitrary and unknown. (In particular they do not have to be Gaussian distributed, or anything like that.) There is a (known, or accessible via samples) distribution $P$ on pairs $(A,B)$ where $A,B \subseteq [n]$. The set $A$ is the set of "revealed" data points -- samples an estimation algorithm has access to. The goal is to estimate some statistic of the data points in $B$, given $A,B$ and the values $x_i$ for $i \in A$. For this review I will just discuss estimating the mean of the values in $B$, although the approach in the paper extends to estimating variance or higher moments as well as to regression problems. Notice that this model allows for situations where the sampling procedure is far from i.i.d. sampling -- in particular, the value $x_i$ of the $i$-th point may have a large effect on whether $i$ is included in $A$ or $B$. The authors study a natural class of estimation strategies for such problems, which they call "semilinear". These are estimators which always output a linear function of the observed values $x$, but where the coefficients of that linear function may depend *arbitrarily* on the realization of $A$ and $B$. This captures standard schemes like importance sampling, as well as a number of interesting additional settings described in the paper. The main result is an efficient algorithm which takes the distribution $P$ (in the form of polynomially-many samples) and outputs an approximately-optimal semilinear estimation strategy (i.e. a rule mapping $A,B$ to a linear estimator, i.e. a weighting scheme for the samples in $A$). The "approximate" in "approximately-optimal" means that the expected in estimating $B$ is at most a constant factor greater than what would be incurred by the optimal semilinear estimator given worst-case data. The authors also extends this result to the linear regression setting. The paper also includes some very interesting numerical experiments in natural settings such as "snowball sampling" where samples are collected via a social network, and demonstrate that their estimator significantly out-performs the sample mean. I think this paper should be of great interest to the NeurIPS community. It checks all the boxes for a great NeurIPS paper: -- focus on a critical problem (how to do statistics in the face of correlated sampling and adversarial data?) -- introduce a new model capturing important existing settings and suggesting new ones -- nice theory result giving an algorithm with strong provable guarantees -- convincing numerical experiments. -- directions for future work

Weaknesses: The paper has few weaknesses. It would be nice if the experiments were for some more realistic settings. The algorithm suggested is probably not practical for more than 50-100 data points, because it requires solving a semidefinite program -- it would be interesting to obtain an algorithm with similar guarantees which scales to larger data sets.

Correctness: I did not check all the proofs line by line, but based on the sketches provided by the authors I think the results are likely to be correct.

Clarity: The paper is well written.

Relation to Prior Work: Is adequately discussed.

Reproducibility: Yes

Additional Feedback: It would be nice to offer some additional argument for lines 270-272.

### Review 2

Summary and Contributions: The paper introduces a new framework for statistical estimation, where we assume knowledge of a distribution which describes the sampling process of the sample set (in machine learning terms, a training set) and the target set, but allows arbitrary values for collected samples (subject to some natural boundedness conditions). Within this general setup the paper provides an efficient algorithm for target mean estimation and its guarantees via a connection to the Grothendieck problem.

Strengths: In my opinion this is a strong submission, especially due to its conceptual novelty. This work arguably falls under the umbrella of robust statistics, however the typical formalization of robust statistics involves some true underlying target distribution, and then the training distribution is obtained as a corrupted version of this target distribution. Here, on the other hand, the assumed distribution is only over the collected samples (in a sense, it’s like assuming a distribution over which samples will be truncated), and then the data values can be arbitrary. The current paper focuses on a relatively simple setting of semilinear schemes, however it’s a very good stab at analyzing this new “randomly collected, worst case data” formalism. I think this submission could inspire a lot of follow-up work.

Weaknesses: I have a couple of stylistic comments. When describing the linear regression setup, I’d add a half-sentence remark clarifying that you’re assuming the model is identifiable (otherwise beta_B is not well-defined). Preceding Definition 2, I was a bit confused by the wording “approximated to arbitrary accuracy”. Given that everything is finite, there is no “approximation”, but you can come up with the appropriate uniform distribution which exactly matches P. Finally, I’m not a fan of the term “sample set”, since technically B is also a sample set (i.e. set of samples). It’s possible that this is some standard terminology though, in which case it’s probably best to stick to whatever is standard. Also, in the future it would probably be good to see what one can say under different constraints of the samples; assuming point-wise boundedness is good when we envision something like i.i.d. samples. However, many natural applications where the samples from A are "coupled" imply, say, an l1 or l2 constraint on the sample set, rather than l-infinity.

Correctness: Yes, the claims seem correct.

Clarity: The paper is very clearly written, I don't have any suggestions regarding this.

Relation to Prior Work: Yes, the Related Work section is thorough.

Reproducibility: Yes

### Review 3

Summary and Contributions: This paper introduces a new model for statistical estimation, which does not assume the data come from a specific distribution. In contrast, the model is based on an assumption on how data are collected. More precisely, for any set of population elements x_1,...,x_n, the estimator draws tuples of sets of indices in [n], (A,B) from a distribution P, where A is the sample set and B is the target set. The estimator only sees the values of x_i, i\in A and aims to give an estimate of the statistic on B. Its performance is measured by the worst case (in terms of population elements x_1,...,x_n) expected error over P. The authors give an algorithm for estimating the target mean (both for the case that P is known and when it is accessed via sampling), and prove its approximation guarantee with respect to the best estimator, for this class of estimators (semi-linear estimators). They also apply their results to the setting of linear regression, giving an estimator for the least-squares regression model of the target set. Finally, the authors demonstrate the better performance of their mean estimator for various sampling processes.

Strengths: -This paper is novel in the sense that it introduces a new model for statistical estimation. This model is based on the assumptions on the sampling process rather than distributional assumptions over the data themselves and it would be interesting if it could yield insights and bounds for cases where distributional assumptions fail (e.g. the important case where data are not drawn iid). The model seems meaningful as it encompasses some of the interesting sampling processes (e.g. snowball sampling, chronological sample structure). -There are interesting mathematical connections between the class of semi-linear estimators and semi-definite programming (that lead to an elegant proof for the approximation factor of the mean estimator).

Weaknesses: -The results are a bit restrictive as they are limited to the class of semi-linear estimators for linear statistics (i.e., sums of functionals of each datapoint in the set). The authors also empirically find that exists a distribution for which the best semi-linear estimator is sub-optimal. However, their structure allows for the clean analysis of their performance (via convex semi-definite optimization). -If we consider n to be the size of the whole population, then the number of samples (tuples of sample/target sets) needed by the estimator, which is poly(n), is quite large. So overall, the results intended to demonstrate the usefulness of the model are not particularly strong. However, this is not necessarily a requirement for a paper introducing a new model.

Correctness: The claims are fully proven and the set-up and results of the experiments are clearly explained.

Clarity: The paper is very well written. I particularly liked the set of three (!) examples aiming to explain why the model is meaningful.

Relation to Prior Work: Prior work mostly focused on removing the iid assumption, but not on removing the distributional assumption completely and was sufficiently discussed.

Reproducibility: Yes

### Review 4

Summary and Contributions: The paper proposes a framework to study a particular kind of subsampling process, where from a finite set two sets are subsampled. The analysis and techniques assume knowledge of the subsampling process but no knowledge of the data distribution is required. The results primarily apply to simple semi-linear estimators which, the authors motivate as being relevant.

Strengths: 1) Novel setting 2) Well written 3) Analysis and principled algorithm provided

Weaknesses: 1) The main assumption of access to P may not hold in many real settings 2) Results mainly apply to scalar sets 3) Results apply to the restricted class of semi-linear estimators, and is not clear how more complex functions could be accomodated

Correctness: Yes, as far as I could verify

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: The paper proposes a framework to study a particular kind of subsampling process, where from a finite set two sets are subsampled. The analysis and techniques assume knowledge of the subsampling process but no knowledge of the data distribution is required. The results primarily apply to simple semi-linear estimators which, the authors motivate as being relevant. I think this is an interesting problem setup where the analysis is based on the subsampling process rather than distribution of the data. The analysis and the algorithm is interesting although primarily restricted to semi-linear models with scalar sets. My main concerns are regarding the realisticness of the setup. The setup assumes subselection from a finite set, which may not be the case for many real machine learning problems. Typically, a bound on n is not known or n is large where the set B is much larger than the set A with it being data collected over time by a deployed system. Another issue is P or the sampling bias is hard to ascertain in many practical circumstances. In some sense, its easier to obtain an estimate for the actual data distribution that the sampling bias as data can come from multiple sources and based on unknown data collection schemes. ---------------------- Post Rebuttal The results for semi-linear models are interesting but some comments on extension to more realistic models would be good for a general NeurIPS reader to have. All in all though given the novelty of the idea I am raising my score.