Global Optimality in Bivariate Gradient-based DAG Learning

Recently, a new class of non-convex optimization problems motivated by the statistical problem of learning an acyclic directed graphical model from data has attracted significant interest. While existing work uses standard first-order optimization schemes to solve this problem, proving the global optimality of such approaches has proven elusive. The difficulty lies in the fact that unlike other non-convex problems in the literature, this problem is not"benign", and possesses multiple spurious solutions that standard approaches can easily get trapped in. In this paper, we prove that a simple path-following optimization scheme globally converges to the global minimum of the population loss in the bivariate setting.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer u4D38/10 · confidence 3/52023-06-22

Summary

The authors give a simple optimization algorithm for DAG-learning-inspired optimization problems that avoids the limitations of known techniques.

Strengths

Originality: Work is original. I particularly liked the reduction from a combinatorial problem to a non-convex optimization one. Quality: Simple and strong paper. Clarity: Clear writing, which can be improved; see questions. Significance: Significant topic/contributions.

Weaknesses

No significant weaknesses found.

Questions

Discussion around Equation (1) is very clean. Line 49: Can you please explain Equation (2) a bit more? Line 64: Is it really easy to see? :) Line 90: Please explain "homotopy." Remark 2: Can you please further explain the difficulties that you mention here? Can you please elaborate on Equation (7)? Line 217: Can you please elaborate on bounding $a$? Please add more details in the caption of Figure 2. Line 235: Why is the interesting regime the one where $\mu < \tau$? Line 258: Not clear! Line 261: "we know" instead of "we known."

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer Y11K7/10 · confidence 3/52023-06-30

Summary

This paper studies the problem of learning the correct Directed Acyclic Graph (DAG) that describes the data using continuous optimization. The connection of this problem with continuous methods comes from a prior work of Zheng et al., where they introduce a differentiable function $h$ whose level set at 0 exactly characterizes DAGs. The problem then becomes to optimize a score function of the data subject to the constraint that h is 0. This is a non-convex problem, since h is non-convex. The standard way of solving it is by converting it to an unconstrained problem and penalizing the solutions with high values of h. This paper provides a way of solving a sequence of these problems with varying penalty parameters, so that the final output provably converges to the true solution, when we have two nodes in the DAG. The model is analyzed in the population case, i.e. when we have access to infinite data. The authors first show that for sufficiently small values of the penalty parameter, the loss landscape is not benign, as there exists a saddle point and a spurious local minimum in addition to the global minimum. Using this observation, they show a way to pick a sequence of penalty parameters, which provides a solution path that avoids the local minima and converges provably to the true model asymptotically. Their main result is phrased in terms of gradient flow, but they also provide a finite step analysis for gradient descent.

Strengths

Consider the optimization problem \begin{equation}\label{eq1} \min_\Theta g_\mu(\Theta) := \mu f(\Theta) + h(W(\Theta)) \end{equation} where $f$ is the least squares loss, $\mu$ is the penalty parameter, $\Theta$ represents the model parameters and $W(\Theta)$ represents the adjacency matrix of the DAG. $f$ is a function that was introduced in prior work, which is $0$ if and only if $W(\Theta)$ corresponds to a DAG. This paper is the first that provides theoretical guarantees about the optimization landscape of this unconstrained problem for various values of $\mu$, in the case of $2$ nodes. It is worth noting that this is actually the method that is used in practice to learn DAGs, since enumeration methods quickly become intractable when the number of nodes in the DAG increases. The authors establish interesting properties of the solutions space, namely that if $\mu$ is sufficiently small, the loss is not benign. Also, the idea of successively solving different optimization problems and using the solution of the last as a starting point for the next is an intriguing one and could have applications beyond this work. The proof strategy is also very intuitive and elegant, by choosing $\mu$ so that the optimizer of $g_\mu$ lies in the basin of attraction of the global minimum of the next problem. All arguments are very clearly explained, which makes this paper enjoyable to read.

Weaknesses

-The result is only proven for $d=2$ nodes in the DAG. This makes most calculations tractable, since there is a closed form for the loss $g_\mu$ (a quadratic polynomial in two variables). Hence, while the proof strategy is elegant and novel, the technical arguments required to establish the claims do not have significant innovation.

Questions

-It would be interesting to see whether similar results hold for higher values of $d$. Have the authors tried to run simulations for $d=3$ or $4$ and observe similar solution landscapes? What about other choices of loss function or penalty? -Is it always true that there is a unique DAG that describes the data? For example, for two nodes, suppose we have $X_2 = X_1 + \epsilon$, where $\epsilon\sim \mathcal{N}(0,1)$. Then, clearly it is also true that $X_1 = X_2 + \epsilon'$, where $\epsilon' \sim \mathcal{N}(0,1)$. In that case, it seems that we have two solutions $x=1,y=0$ and $x=0,y=1$. I'm curious where would this come up in the analysis.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Yes

Reviewer xJpx5/10 · confidence 1/52023-07-15

Summary

This paper presents a novel approach to the non-convex optimization problems associated with learning the structure of a structural equation model (SEM) or Bayesian network. Considering the equivalent penalty form, the authors propose a homotopy-based optimization scheme that finds global minimizers of the problem by iteratively decreasing the penalty coefficient according to a given schedule. They prove that this algorithm converges globally to the global minimum, regardless of the initialization for W. The authors also demonstrate that the non-convex program is non-benign, meaning that naïve implementation of black-box solvers are likely to get trapped in a bad local minimum.

Strengths

I am not an expert in this field. But the paper's findings seems to have significant implications in the context of learning the structure of SEMs or Bayesian networks. The paper is well-structured and the authors clearly explain their methodology and findings. They also provide a clear visualization of the non-convex landscape and the solution trajectory.

Weaknesses

The authors' approach is primarily focused on the bivariate case, which may limit its applicability in more complex settings. Some numerical study on large-scale problems is desirable.

Questions

How does the landscape look like in high dimensional setting? Can you provide any experiments to more than two variables?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Does not apply

Reviewer NK3q5/10 · confidence 3/52023-07-31

Summary

the paper provides a theoretical study on the loss landscape and convergence in gradient-based DAG learning framework. By focusing on linear functions with the number of variable d = 2, They provide a homotopy-based optimization scheme to guarantee the global optimality. Some numerical validations are provided.

Strengths

the paper focuses on the theoretical study on the global optimality and convergence. This is an important but difficult questions. The theoretical results mainly shows the initialization regime as requirements for convergence, which makes sense given the nonconvex nature, although I did not check the proof in details (given its length). Illustrative examples and results are provided.

Weaknesses

Unfortunately, the current results are only applicable to d=2 and linear functions (understandably, of course). No discussions on how nonlinear or a large number of variables would affect the loss landscape, and/or how the homotopy algorithm would be affected. Would be more interesting if the insights of the theoretical results can be used in practical algorithms.

Questions

- "...bears a resemblance to the success of training deep models, which started with AlexNet for image classification": First of all, modern deep models have success before alexnet on imagenet, so this is not accurate. Second, such a statement is over the top and should be removed. - L82: indeed only two DAGs are in the entire search space yet the analysis is quite complex. There is no indication on how such analysis and approach would scale or behave with an increasing number of variables, hence it is hard to judge the usefulness of these analysis yet. In comparison, although the combinatorial search is not used, its expected performance is the same with larger d. - Eq 4: x is already used to represent data, and should avoid using it as parameter (lack of notation clarity). - Eq 5: while re-ordering variable to obtain such a upper triangular structure is not an issue here yet, in datasets W* would generally not be so. How does a non-upper triangular structure impact analysis, for example a's usage afterward? - Some derivation are not given fully. For example, f(W) and h(W) skips many steps in between. - L199: what is k here? - the regime of initialization for \mu: very interesting observation. - optimization landscape with nonlinear functions: current analysis focuses on linear function. How would it change for nonlinear functions (such neural networks), even for d=2? Given the nonlinear nature, one can imagine it would make the analysis even hard, given two layers of neural networks are the most complex cases that has some theoretical understanding. - can the homotopy-based optimization scheme be used to improve convergence of NOTEARS related algorithms?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

not discussed but not needed.

Reviewer xJpx2023-08-19

Thank you for the clarification. There could be some interesting future directions to pursue, but my concern about the limitation of this current paper is not fully addressed. I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC