NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:4413
Title:Sparse Logistic Regression Learns All Discrete Pairwise Graphical Models

Reviewer 1

This paper gives a simple and elegant algorithm for solving the long-studied problem of graphical model estimation (at least, in the case of pairwise MRFs, which includes the classic Ising model). The method uses a form of constrained logistic regression, which in retrospect, feels like the "right" way to solve this problem. The algorithm simply runs this constrained logistic regression method to learn the outgoing edges attached to each node. The proof is elegant and modular: first, based on standard generalization bounds, a sufficient number of samples allows minimization of the logistic loss function. Second, this loss is related to another loss function (the sigmoid of the inner product of the parameter vector with a sample from the distribution). Finally, using a result from Klivans and Meka, they relate this function to the true error in the parameter vector. The authors also show that approximate minimization of the loss function can be done computationally efficiently via mirror descent. The sample complexity improves (slightly) upon the previous best results in case where the alphabet is k-ary (reducing the dependence on k from k^5 to k^4), but this is not the reason why I like the result. The thing I like most about this work is that it proposes and analyzes such a simple algorithm for this long-studied problem. I believe this method should be extendible and applicable to more general settings, due largely to the simplicity of its core. Aside from this, the "right" answer for such a long-studied problem definitely deserves publication. This is perhaps the main contribution -- showing that logistic regression can be used as the main primitive, rather than the more complicated Sparsitron of Klivans and Meka. The modular nature of the proof greatly helps with the clarity. This is a fairly short review, but I don't have much more to say: it's a nice result which I believe should be accepted. If I had to criticize the paper, I would say that the theoretical sample complexity is not significantly improved over previously-known methods. That said, as made clear above, I think the method itself warrants publication. It would be interesting to see a more thorough empirical evaluation, to compare with the interaction screening method and in more settings than just the very basic ones studied here. But I don't think this evaluation is necessary for acceptance.

Reviewer 2

I read your feedback. I understood that your main contribution is theoretical (and always thought it is interesting). I increased my overall score though still believe that within the eight page limit your paper could benefit from focusing on the Ising case and only briefly mention the extension to larger alphabets. But of course I know that this means a major change that is probably not feasible. This paper makes an interesting contribution. The problem formulation as l1-regularized logistic regression (in the binary case) and regularized l(2,1)-regularized softmax regression problem (in the general) is, as the authors also acknowledge, well known. The consistency (parametric and structural/sparsistency) analysis is new and improves in certain (small) aspects over previous work. Formally, the improvement over previous work is a slightly better sample complexity (the dependency on the size k of the sample spaces of the variables is only k^4 instead the previously known bound of k^5). Unfortunately, I am lacking the time to check the technical soundness of the paper. All proofs are given in an appendix that I did not read. The presentation, although the proofs are deferred to an appendix, is quite technical. The given outline of the proof is too short (at least for me) to get the underlying ideas. Experiments are only presented for rather small examples (up to 14 variables, up to k=6). Still, it is demonstrated the presented approach does indeed not need the incoherence assumption.

Reviewer 3

This work tackles the problem of learning the edge weights in general discrete pairwise MRFs, of which the classical Ising model is a particular instance (when the nodes have cardinality k = 2). A steady body of work has recently addressed this problem and made strides towards structure learning with efficient sample and time complexity. The paper contributes algorithms to learn both the structure of Ising models (using l_1 constrained logistic regression) and the general case (k > 2, using l_{2,1} constrained logistic regression). These algorithms are theoretically shown to improve upon the sample complexity of the state-of-the-art. Further theoretical results show that, when the presented mirror descent algorithm is used during optimization, the time complexity of learning may be made competitive with the state-of-the-art without sacrificing statistical guarantees. The theoretical results are supported with several experiments, which soundly show that the presented algorithms outperform the state-of-the-art in sample complexity and work with fewer model assumptions than other previous works. The paper is well written and the results contribute steady progress towards more efficiently solving this particular problem. As expected, the paper is very theoretical, but the authors did an excellent job of both keeping the main paper readable while doing further heavy lifting in the supplementary. Furthermore, the extra content in the supplementary was very helpful as a whole (e.g., Section A was very comprehensive, the extra experiments were appreciated, and the breakdown of the proofs was well done).