Optimizing over trained GNNs via symmetry breaking

Optimization over trained machine learning models has applications including: verification, minimizing neural acquisition functions, and integrating a trained surrogate into a larger decision-making problem. This paper formulates and solves optimization problems constrained by trained graph neural networks (GNNs). To circumvent the symmetry issue caused by graph isomorphism, we propose two types of symmetry-breaking constraints: one indexing a node 0 and one indexing the remaining nodes by lexicographically ordering their neighbor sets. To guarantee that adding these constraints will not remove all symmetric solutions, we construct a graph indexing algorithm and prove that the resulting graph indexing satisfies the proposed symmetry-breaking constraints. For the classical GNN architectures considered in this paper, optimizing over a GNN with a fixed graph is equivalent to optimizing over a dense neural network. Thus, we study the case where the input graph is not fixed, implying that each edge is a decision variable, and develop two mixed-integer optimization formulations. To test our symmetry-breaking strategies and optimization formulations, we consider an application in molecular design.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer eLVB6/10 · confidence 2/52023-06-18

Summary

To overcome the symmetry issue when solving inverse problems on trained GNNs, this paper proposes two types of symmetry-breaking constraints to break symmetry. The authors construct an indexing algorithm, and prove that the resulting graph indexing satisfies the proposed symmetry-breaking constraints. They also develop two mixed-integer optimization formulations in the case where the input graph is not fixed.

Strengths

1. The symmetry issue which this paper aims to overcome is important, and the proposed indexing is a general approach that applies in many kinds of situations. 2. The theoretical analysis is solid, and the proof seems to be correct.

Weaknesses

1. The experiments are not persuasive enough. The reasons are as follows. (1) For Sec 3.1 (Mixed-integer optimization formulation for molecular design): In my opinion, a formulation itself WITHOUT experimental verification should not be thought of as a contribution. (2) For Sec 3.2 (Counting feasible structures: performance of symmetry breaking): The experimental results (Table 1) do not have baselines. 2. There is no related work analysis (except a little description in Introduction). That makes it a little hard to justify the originality. For example, “The second formulation generalizes the big-M formulation for GraphSAGE in [70]” (L197). However, the authors did not explain how the big-M formulation looks like, in which way the proposed formulation generalizes it, and are there overlaps between the constraints of the big-M formulation and the proposed ones?

Questions

1. I suggest that the authors provide complexity analysis to the proposed algorithm, since in Table 1 there are many time out cases. 2. I suggest that the authors discuss the relationship between the proposed algorithms and related works (please see Weaknesses (2)). 3. I suggest that the authors provide necessary experiments (please see Weaknesses (1)).

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

3 good

Limitations

Yes.

Reviewer 5u3v7/10 · confidence 3/52023-06-24

Summary

This paper investigates the optimization of trained GNNs. This is a permutation-invariant problem and all points on an orbit of the symmetric group have the same performance. The authors proposed some symmetry-breaking approaches to have smaller search space and hence less redundancy and more efficient optimization. The correctness is proved theoretically and the efficiency is examined numerically.

Strengths

1. Optimizing a trained GNN is an important topic in practice and breaking symmetry can reduce the redundancy in many optimization problems. 2. The authors provide the theoretical guarantee of the proposed algorithm. 3. The numerical results look nice. It is reasonable to expect such improvement since the MIP formulation does have large symmetry group.

Weaknesses

1. I do not think this paper has enough literature review on detecting and breaking MIP symmetry -- how do previous work detect symmetry? what is the cost for symmetry detection? what is the connection/difference between previous symmetry-breaking approaches and the proposed approach? I think Gurobi can detect symmetry but the computational cost might depend on the size of the symmetry group. For optimizing GNNs, the symmetry group is obvious (just the permutation group). So one does not need to detect the symmetry and can design problem-specific symmetry-breaking approaches. Those ideas/background should be discussed more in the paper. 2. As the authors discuss in Section 4, the numerical experiments are not general enough, but I think the current experiments have already shown that the approach is promising.

Questions

If you have Gurobi to detect and break the symmetry, will Gurobi add any constraints and if so, are those constraints the same as or different from (C26) and (C27)?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors clearly discuss the limitations in Section 4 and I agree with their discussion.

Reviewer ni197/10 · confidence 3/52023-07-06

Summary

This paper suggests using a symmetry breaking indexing for nodes in a graph. The argument is that for inverse problems, such as molecule design, graph isomorphism results in finding many equivalent solutions. The symmetry breaking indices are supposed to partially alleviate this redundancy. They use pre-trained GNN (based on GraphSage) for computer-aided molecule design (CAMD). During the design, they use Mixed-Integer Programming (MIP) (solved using Gurobi) to solve a set of constraints to make the molecules chemically realistic as well as constraints arising from their indexing scheme. They conduct experiments on QM7 and QM9 and find that the symmetry breaking methods perform much better. They show more successful runs (i.e. producing molecules satisfying the constraints) and achieve this in a fraction of the time, compared to models without symmetry breaking.

Strengths

The paper makes a good case for how adding an ordering (indexing nodes in a specific manner) can dramatically improve inverse problems on graphs, e.g. designing molecules, solved using MIP. The paper is well structured and discusses the literature in good depth. It explains the algorithm and experimental setup. The results are also very strong for the experiments.

Weaknesses

The experiments show that with the indexing, MIP converges much faster. Table 1 states that MIP find many feasible solutions for each setting and set of constraints. One thing I hope the authors can assess (maybe using a Weisfeiler-Lehman hash or similar; see questions below) is how much the symmetry breaking *actually* removes isomorphic solutions, compared to the runs without symmetry breaking. they could do it for a small setting, maybe QM9, N=3 or 4. The goal would be two fold: 1. To show that the symmetry breaking indeed removed many isomorphic molecules 2. to check how the diversity of solutions in the symmetry broken case compares with the diversity of molecule found in the vanilla algorithm. I was also hoping some kind of baseline could be discussed, but I understand that to have a fair comparison they need to use the same pretrained GNNs with and without symmetry breaking. Also, the claims about generality of the approach and applicability to other graph problems, though reasonable, lack any substantive evidence or hint as to how to do it concretely. I suggest wither softening the claims or giving some hints as to why that is. In terms of structure, I think the proof of theorem 1 is not necessary to be in the main text and occupies too much space. Instead some version of the last two tables in the appendix could be more useful in the main text.

Questions

1. Table 1: In each row, is higher necessarily better? If we want less symmetry copies, should sometimes less be better? 2. How do we know that the solutions found with the symmetry broken model are as diverse as the one with symmetries? These are small molecules. Can you use a fast Weisfeiler-Lehman type hash to check isomorphism of found molecules at least in small N experiments? Maybe also use a measure of diversity, some kind of entropy or something. ### Minor questions: line 75 “Directed to undirected” how? line 106 A is now a multiset, not the adjacency matrix anymore? line 183 (*) is not permutation invariant, as written. $w_{u\to v}$ is like a generalized Graph Attention (GAT). line 187 Dense NN: do you mean a fully connected layer (no graph)?

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors discuss technical limitations, but I don't think there are immediate societal impacts, as the paper pertains to a technical method to speed up generative design in graph problems.

Reviewer eveB7/10 · confidence 3/52023-07-07

Summary

This paper presents novel symmetry-breaking constraints for optimizing trained graph neural networks (GNNs) and addresses the graph isomorphism issue. The authors develop two mixed-integer optimization formulations and evaluate their methods in the context of molecular design.

Strengths

1. **Relevance and Application**: The paper addresses the important issue of GNN optimization and demonstrates practical utility through molecular design application. 2. **Empirical Evidence**: The authors provide convincing experimental results supporting their claims, underscoring the proposed method's effectiveness. 3. **Clarity**: The manuscript is well-written, presenting complex ideas with notable clarity.

Weaknesses

1. **Insufficient exploration of S1 and S3 compatibility**: The paper falls short in exploring whether S1 and S3 requirements can coexist for certain indexings of a connected graph with features. 2. **Insufficient explanation of structure feasibility constraints**: There is a lack of clarity concerning the interplay between structure feasibility constraints C1 - C21 and the indexing process. Specifically, the paper does not elucidate the number of these constraints that are related to indexing and if they are compatible with S3.

Questions

1. Do the indexing produced by Algorithm 1 for a connected graph satisfy S1?

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

n/a

Reviewer eLVB2023-08-12

Response to Authors

Thanks for your reply. You have already solved my concerns, so that I improve the score.

Reviewer 5u3v2023-08-13

Thank you very much for your reply. The discussion about the previous literature and connection/difference look very nice.

Reviewer eveB2023-08-16

Thank you for your response. Could you please provide a full proof for the proposed Lemma 2?

Authorsrebuttal2023-08-16

Please see the full proof as follows: **Proof of Lemma 2** Given any connected graph $G=(V,E)$ with any indexing that satisfies (S3), we prove that this indexing also satisfies (S1) by induction. Node $0$ itself is a connected graph. Assume that the induced subgraph consists of nodes $0,1,\dots, v$ is connected, it suffices to show that the induced subgraph consists of nodes $0,1,\dots, v+1$ is connected. Equivalently, we need to prove that there exists $u<v+1$ such that $A_{u,v+1}=1$. Assume that $A_{u,v+1}=0,\forall u<v+1$. Since the graph is connected, there exists $v'>v+1$ such that: $$ \begin{equation*} \begin{aligned} \exists u<v+1,s.t.A_{u,v'}=1 \end{aligned} \end{equation*} $$ Then we know: $$ \begin{equation*} \begin{aligned} \mathcal N(v+1)\cap [v+1]=\emptyset, \mathcal N(v')\cap [v+1]\neq \emptyset \end{aligned} \end{equation*} $$ Recall the definition of $LO(\cdot)$, we have: $$ \begin{equation} \begin{aligned} LO(\mathcal N(v+1)\cap [v+1])>LO(\mathcal N(v')\cap [v+1]) \end{aligned} \end{equation} $$ Since the indexing satisfies (S3), then we have: $$ \begin{equation*} \begin{aligned} LO(\mathcal N(w)\backslash\{w+1\})\le LO(\mathcal N(w+1)\backslash \{w\}),\forall v<w<v' \end{aligned} \end{equation*} $$ Applying Property 3 gives: $$ \begin{equation*} \begin{aligned} LO((\mathcal N(w)\backslash \{w+1\})\cap [v+1])\le LO((\mathcal N(w+1)\backslash \{w\})\cap [v+1]),\forall v<w<v' \end{aligned} \end{equation*} $$ Note that $w>v$. Therefore, $$ \begin{equation*} \begin{aligned} LO(\mathcal N(w)\cap [v+1])\le LO(\mathcal N(w+1)\cap [v+1]),\forall v<w<v' \end{aligned} \end{equation*} $$ which means that: $$ \begin{equation} \begin{aligned} LO(\mathcal N(v+1)\cap [v+1])\le LO(\mathcal N(v')\cap [v+1]) \end{aligned} \end{equation} $$ The contradiction between $LO(\mathcal N(v+1)\cap [v+1])$ and $LO(\mathcal N(v')\cap [v+1])$ completes the proof.

Reviewer eveB2023-08-16

Thank you for your reply! Can you elaborate on the derivation of the following step: Note that $w>v$. Therefore, \begin{equation*} L O(\mathcal{N}(w) \cap[v+1]) \leq L O(\mathcal{N}(w+1) \cap[v+1]), \forall v<w<v^{\prime} \end{equation*} Specifically, are you using $L O(\mathcal{N}(w+1) \cap[v+1])= L O((\mathcal{N}(w+1) \backslash w) \cap[v+1])$? If so, it seems the corner case where $w = v+1$ needs additional details.

Reviewer eveB2023-08-16

I see; thank you for the clarification. I am happy to raise the rating to 7.

Reviewer ni192023-08-16

Thank you

Thank you for the clarifications about Table 1 and the rest of the paper. The diversity comment also makes sense, I think.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC