Interpretable Graph Networks Formulate Universal Algebra Conjectures

The rise of Artificial Intelligence (AI) recently empowered researchers to investigate hard mathematical problems which eluded traditional approaches for decades. Yet, the use of AI in Universal Algebra (UA) -- one of the fields laying the foundations of modern mathematics -- is still completely unexplored. This work proposes the first use of AI to investigate UA's conjectures with an equivalent equational and topological characterization. While topological representations would enable the analysis of such properties using graph neural networks, the limited transparency and brittle explainability of these models hinder their straightforward use to empirically validate existing conjectures or to formulate new ones. To bridge these gaps, we propose a general algorithm generating AI-ready datasets based on UA's conjectures, and introduce a novel neural layer to build fully interpretable graph networks. The results of our experiments demonstrate that interpretable graph networks: (i) enhance interpretability without sacrificing task accuracy, (ii) strongly generalize when predicting universal algebra's properties, (iii) generate simple explanations that empirically validate existing conjectures, and (iv) identify subgraphs suggesting the formulation of novel conjectures.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Peer review

Reviewer swnT6/10 · confidence 3/52023-06-27

Summary

This paper introduces the novel application of Artificial Intelligence (AI) in the field of Universal Algebra (UA), which has remained unexplored until now. The authors propose a methodology that combines equational and topological characterizations with graph neural networks to investigate UA conjectures. To address the limitations of transparency and explainability in graph neural networks, they develop an algorithm for generating AI-ready datasets based on UA conjectures and introduce a new interpretable graph network layer. Experimental results demonstrate that interpretable graph networks offer enhanced interpretability without compromising task accuracy, successfully validate existing conjectures, and identify subgraphs that suggest the formulation of new conjectures. Overall, this paper presents a pioneering approach that integrates AI with UA, enabling empirical validation and formulation of conjectures in this foundational field of mathematics.

Strengths

This paper propose to use GNN to study the UA subject in the modern mathematics. The paper claims to have generated the first UA dataset, which is a previously untouched area. This paper also proposed the iGNN which improves model interperability and identifies the relevant subgraphs.

Weaknesses

As the authors pointed out, the method only applies to finite number of lattices.

Questions

na

Rating

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

Limited to finite lattices: The methodology is restricted to finite lattices, which may not capture all the relevant information about infinite algebraic structures. While insights gained from finite-lattice explanations can still provide valuable information, the generalization to infinite structures may be limited. Restriction to topological properties: The approach focuses on topological properties on graphs, and it may not be applicable to non-structural properties that require the use of other interpretable models. While the method is effective in characterizing equational and topological conjectures, it may not cover a broader range of mathematical properties that are not solely determined by the graph structure.

Reviewer 3uvy4/10 · confidence 3/52023-07-06

Summary

In this paper, the authors explored the use of AI in mathematical problems Universal Algebra. They propose an algorithm for generating datasets of lattice varieties and derive an interpretable graph neural networks for predicting universal algebra’s properties. Experiment results show the effectiveness of the proposed method. Overall, the motivation is clear and the paper is well written.

Strengths

The problem the paper is investigating is of interest and importance. The authors propose a relatively complete pipeline consisting of both datasets generation and a graph neural network for interpretable prediction. The paper is well written and the problem definition is clear. In addition, the authors provide examples to emphasize how the Gumbel softmax works which makes it more readable. Experiment results seem reasonable. It is good that the authors verify their method on datasets with different settings and use qualitative metrics for interprebility.

Weaknesses

One concern is from the dataset generation. As the authors mentioned that the number of existing lattices increases exponentially as n increases, I am curious whether the algorithm is really scalable. In addition, it seems that the algorithm is unable to control label distribution generated. Using Gumbel softmax instead of regular activation functions seems novel. However, one concern is whether the semantic concept acquired via the one-hot embedding is really meaningful. Because two embeddings may not necessarily have correlation when they have the maximum value in the same indice. The authors mainly compare their method with black-box GNNs and show better results. However, no recent works are included for comparison. Though the authors discussed the difference between the proposed method with existing general graph explanation methods. There are actually some ad-hoc explanation methods that do not rely on subgraph patterns, e.g. GSAT. Ref: Miao S, Liu M, Li P. Interpretable and generalizable graph learning via stochastic attention

Questions

What is the running time for generating the datasets used in the experiments? How is the time increased with n? Can the authors include recent general graph explanation methods for comparison?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

3 good

Limitations

It is recommended that the authors can provide results in a table to be more readable.

Reviewer LfMh5/10 · confidence 2/52023-07-07

Summary

This paper aims to use AI as a tool to help solve problems in mathematics. In particular, this paper focuses on universal algebra, which studies algebraic structures and can be considered as learning properties of graphs. The contributions are two folds: - This work proposes a method for generating a benchmark for universal algebra, and releases a dataset of 29000 lattices (partial orderings, which can be represented as graphs) of different sizes and 5 properties. - Generation: filter binary matrices (each lattice on a set of size n can be represented by a nxn matrix) that admits a partial ordering. - This work proposes an interpretable neural layer (iGNN) that can be used to solve tasks in the benchmark dataset. Empirical results show that the model is able to generalize across lattice sizes while being interpretable.

Strengths

- The paper provides the required background and is self-contained. - The proposed benchmark dataset is the first dataset for universal algebra which is valuable to the field. - The proposed interpretable layer achieves good empirical performance. - Experiments in Section 5.1 show that on UA tasks, the proposed iGNN scores higher in terms of interpretability metrics, while preserving the performance of black-box GNNs. - The interpretability results in Section 5.2 are interesting: the appearance of substructures corresponding to N5 and M3 is consistent with the theoretical understanding.

Weaknesses

- The dataset generation process is relatively straightforward (i.e. filtering to get valid lattices, and sampling since it's computationally infeasible to enumerate all configurations). - Update post-rebuttal: this is no longer a concern: the author responses have discussed more details and considerations on scalability. - The proposed neural layer is interpretable due to constraining the function classes, which may limit the representational power of the network. - For the pooling layer, the paper chooses max or sum pooling, which reflects the existence or number of 1s in the pooled inputs; this seems to provide only a limited level of interpretability and I wonder if other aspects can be used.

Questions

- Sec 3.2.1, node level features: the "vice versa" part is unclear: we only have the converse direction if we assume that the mapping has some reverse Lipschitz condition (e.g. if the neighborhoods are distinct, then the corresponding $h_i$'s are different), which is not guaranteed in the current setup? - top of page 5: where do the values of $h_{II}$ come from? Are the values made up as examples? - Sec 3.2.3: what are examples of intermediate loss functions? How to obtain the signals/labels for the loss? - Sec 5.2: what does a "concept" mean? How are the weights computed? Is there a set of "concepts" (small graphs?) to search over? Writing related: - A notational question: why $\approx$ rather than =? - The background can be defined more clearly in the main text (or please consider adding pointers to the appendix): - explain that $\vee, \wedge$ correspond respectively to join (least upper bound) and meet (greatest lower bound). - explain $M_3$ and $N_5$ are named for "Modular" and "Nonmodular". - line 77: is there a typo? - to help understand the properties, it might be helpful to point out that not modular implies not distributive. - Sec 3.1: are $n,L$ the same thing? - Fig 6: where's $M_3$?

Rating

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

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

2 fair

Limitations

The paper discusses limitations that the proposed methods are constrained to finite-size graph and only work with structural properties. There is no direct societal impact.

Reviewer AeQ96/10 · confidence 3/52023-07-07

Summary

A general methodology is introduced that allows a researcher interested in problems regarding lattices, in the mathematical field of Universal Algebra, to investigate any algebraic property whose validity can be verified on a finite lattice. The paper indicates how a dataset can automatically be generated . Then a new GNN variant is devised, adapted from previous works (e.g. [1]), that is interpretable. This model is compared to the performance of a vanilla, black-box GNN, with all the other model parameters kept identical. [1] Steve Azzolin, Antonio Longa, Pietro Barbiero, Pietro Liò, and Andrea Passerini. Global ex- plainability of gnns via logic combination of learned concepts

Strengths

The paper discusses how a field that has hitherto not seen a large number of AI applications be treated. Two new GNN variants are proposed, called iGNN and HiGNN. I also like the fact that the well-known lattices N_5 and M_3 where used throughout the text as examples of how the learning algorithms works (Example 3.1 and 3.2).

Weaknesses

__I. Various hyperbolizing adjectives are used throughout the text, that do not provide information:__ - line 18: "Universal Algebra (UA, (6)) is one of the foundational fields of modern mathematics" The importance of Universal Algebra is somewhat overstated. While it indeed is a foundational field, it is usually not part of a typical university curriculum. - line 43 onwards: "First, we propose a novel algorithm that generates a dataset suitable for training AI models based on an UA equational conjecture. Second, we generate and release the first-ever universal algebra’s dataset compatible with AI [...] Our findings demonstrate the potential of our methodology and open the doors of universal algebra to AI." The relevance of the contribution is hyperbolized and overstated, for the following reasons: 1) Indeed, this may be the first application of AI to universal algebra (although experience has shown that in machine learning often new results are introduced, that under a different name were already known previously -- --, so unless an extremely thorough literature review would have been made, I would not feel confident making this assertion). 2) It is known that AI is very well suitable to a particular type of mathematics, where the involved objects are discrete (see [1]), so it is not surprising that this approach works well in this setting too. 3) I am not convinced paper does not "open doors of universal algebra to AI" (see also the claim repeated on line 408), since the method is not tied to the field of Universal Algebra, but rather to the objects of study. Universal Algebra has strong connections to model theory, for example (see [2], chapter 5, which is a reference that is cited in the paper), and exploring these using machine-learning is definitely not covered under the current approach. line 90: "advanced AI methods" (what exactly makes an AI method "advanced" is debatable) line 396: "However, as universal algebra is a foundational branch of modern mathematics, any contribution to this field can already have significant implications in various mathematical disciplines." I feel like these hyped-up claims obscure the actual strong points of the paper: A well-executed experiment with a suitable neural model (GNN) (and other strong points outlined in the Strength section). __II. Other observations__ - while not a weakness, but a curiosity: it is noteworthy that the lattice examples from Figure 2, page 2, are identical to those from the Wikipedia article on lattices, see Pic 10 and Pic 10 [3]. In Wikipedia, it is mentioned that these are the smallest lattices of their type, but the paper does not mention this (tangential) fact. - in the appendix, a number of definitions from Universal Algebra are taken from well-knowns, e.g., Definition A.3 corresponds ad literam to Definition 9.3 from [2]. I feel it could be emphasized more strongly that many of the definitions are directly copied from existing sources, rather than merely stating " More formally (6; 8; 16):", where their reference 6 is my reference [2] below. More precisely, at the beginning of the appendix, it is stated: "An n-ary operation on A is a function that takes n elements of A and returns a single element from the set. More formally (6; 8; 16):" which implies that only the first ensuing definition of arity is taken from references rather than all ensuing definitions. It would be helpful, to either explicitly state that all theorems/definitions are taken from books (and ideally, like in Theorem A.5, show from where the theorem comes). - on table from the appendix is missing (below line 582): "Note: ADD TABLE WITH CONCEPT PURITY". This should be fixed, if the paper is accepted, but doesn't make a difference for acceptance/rejection decision. [1] A.Z. Wagner, Constructions in combinatorics via neural networks, https://arxiv.org/abs/2104.14516 [2] Burris S., Sankappanavar H.P., A Course in Universal Algebra, Springer Millenium Edition [3] https://en.wikipedia.org/wiki/Lattice_(order)

Questions

- Is there a reason you chose to use linear classifiers instead of decision trees, as the interpretable model of choice? - consider line 150 ("To address this, we use a hard Gumbel-Softmax activation") vs. line 219 ("The choice of the activation and loss functions to train iGNNs depends on the nature of the task at hand and does not affect their interpretability"). Does that mean that you use different activation functions at different parts of the GNN?

Rating

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

2 fair

Limitations

The problem that is solved is quite narrow, though the authors go out of their way to make sure that this is not emphasized.

Reviewer AeQ92023-08-18

I thank the authors for their response, in particular the elucidations to the last two questions (I could recommend adding them to the main paper). > Secondly, the field commonly relies on deterministic algorithms, utilized to construct finite models or serve as deterministic theorem provers, such as "mace4" and "prover9" [1]. In this sense, we hope that our paper could represent a guidance to further explore the study of UA’s problems with AI. I'm intrigued by this remark. Could you add more motivation how your method might interact with (or even improve) these theorem provers? i am not too familiar with them, so maybe I am mistaken when I am saying this, but it seems to me that these provers use finite models to create counterexamples to claims?

Authorsrebuttal2023-08-20

We extend our gratitude to the Reviewer for the appreciation of our response and for engendering a thought-provoking discussion. “*the elucidations to the last two questions (I could recommend adding them to the main paper).*” In accordance with the Reviewer's suggestion, we intend to add two additional sentences within the revised version of the paper, clarifying the central aspects of the two questions. “*I'm intrigued by this remark. Could you add more motivation how your method might interact with (or even improve) these theorem provers? i am not too familiar with them, so maybe I am mistaken when I am saying this, but it seems to me that these provers use finite models to create counterexamples to claims?*” “*mace4*” is a “model-maker” that is used mainly to investigate finite models with a certain axiomatisation. The Reviewer correctly identified the main utility of this tool in searching counterexamples, however it is also used in various other tasks. For instance, it can be useful also to improve the comprehension of the structures that model specific axioms and to understand how they change removing some of the axioms or adding others (i.e. which of them are not satisfying certain additional axioms?, removing one of the axioms which models are included?, and so on). “*prover9*” is instead a classic theorem-prover that, given a set of axioms and a statement to prove, outputs a proof of the claim if it is able to find it. Both of these tools are quite known in algebra and they can be integrated by our framework which provides an orthogonal information w.r.t to the ones given by “*prover9-mace4*”. Indeed, the process of discovery in mathematical research involves a loop of the following steps: 1- Having an idea about a mathematical property. 2- Conjecturing a statement. 3- Either prove the statement or find a counterexample. 4- Refine the idea and the conjecture, if not proved. Clearly, “*prover9-mace4*” are mainly used for step 3 of this loop, while our framework aims to help phase 1-2 (usually the most difficult parts) providing a better understanding of structures and eventually leading to new ideas. To illustrate, let us suppose for example that for some reason we are interested in introducing and studying new mathematical structures characterized by a set of axioms A (a common problem in algebra, logic, and other fields). Then, *mace4* can be used to conduct a preliminary investigation, to understand how the finite models of A look like for small cardinalities usually (in the case of Algebra the first questions are: How many finite models of small cardinality there are? Which are the congruences of these structures?, Can they be characterized in a specific way?, Which are the subdirectly irreducible structures?, and so on). Then, our framework could be applied to the congruence lattices of those structures, which often encode many relevant properties of the structures, looking for recurrent patterns that can characterize these properties (as a new conjecture). In the last part *prover9* and *mace4* could be then used to prove or disprove the conjectured statement. The authors once again express profound gratitude for the keen interest demonstrated by the Reviewer and for the highly stimulating discourse that has emerged from her/his insightful observations.

Reviewer AeQ92023-08-21

Unfortunately, there is not enough time now; but showing how your work could be integrated in a loop with prover9 and mace4 seems to me to really be the biggest original contribution (otherwise it is just a technical improvement over the current state of the art); I will leave my score as it is for now, but perhaps you will be able to find time to make this update for the camera ready version. At least including an explanation in the paper, as you did above, would be highly beneficial for readers.

Authorsrebuttal2023-08-21

Thank you once again for your feedback, we will include a summary of this discussion in the camera ready. We do agree that the connection with existing provers could be an interesting way forward of practical use for algebrists.

Reviewer LfMh2023-08-20

Apologies for my late reply! Thank you very much for the clarifications; I have raised my score. May I have some further clarifications please: - About limiting the representation power: the current experiments do show that the representation power of iGNN is sufficient for solving these UA tasks, but I was wondering whether/how iGNN would be limiting in general: could you comment on what tasks would black-box GNNs be able to solve but iGNN would not? How is this related to the observation that black-box GNN outperforms iGNN on smaller number of nodes? - Clarification about strong vs weak generalization: is the "strong vs weak" distinction defined somewhere?

Authorsrebuttal2023-08-21

Thank you for raising your score and engaging the discussion! We know it’s not the ideal time of the year to serve as a reviewer. Here we provide further clarifications on the raised questions. **About limitation of the representation power** The representation power of iGNNs is generally constrained by the size of the concept embeddings $q$, unlike standard GNNs. In practice, the Boolean nature of concept embeddings limits the range of representable concepts to $2^q$, rendering iGNNs unable to perfectly handle tasks requiring more than $2^q$ concepts. Additionally, the Gumbel Softmax's enforcement of Boolean representation leads to more rigid and bumpy training, especially when working with datasets containing limited informative samples, as seen in small graphs in our experiments. Also in our experiments we used an interpretable classifier based on a linear layer because it was enough to solve the tasks. However, for more challenging tasks (e.g.,dependent on a non-linear combination of concepts) other interpretable and differentiable models could be considered, without the need to change the backbone structure of iGNNs. **Strong vs weak generalization definition** We draw the distinction between "strong vs weak" generalization based on [1], which uses the term "strong generalization" to describe GNN models' performance on larger graphs than their training set. This type of experiment falls under out-of-distribution testing, a highly challenging scenario for statistical learning models. We will add this reference to the main paper when introducing the two tasks. [1] Veličković, Petar, et al. "Neural execution of graph algorithms." arXiv preprint arXiv:1910.10593 (2019).

Reviewer LfMh2023-08-21

(I guess there is never an "ideal" time to serve as a reviewer, so it really is my fault rather than the timing, but thank you for your understanding :)) Thank you for these further clarifications!

Reviewer 3uvy2023-08-21

Thanks for the authors' effort on discussing my questions. Most of my questions have been addressed. However, from the comparison between iGNN and GSAT, they achieve comparable performance and GSAT is also ad-hoc mechanism. The novelty of the proposed method is undermined. I would like to keep my current ratings.

Authorsrebuttal2023-08-21

“*The novelty of the proposed method is undermined*”. First, we would like to highlight that our novel architecture is one of the three contributions presented in the paper. In any case, we do not agree with the Reviewer for 3 reasons: 1- **Novelty in results**: our results show that iGNNs do attain better results than GSAT on one of the key metrics: *The completeness score of iGNNs in strong generalization mode is better than GSAT by up to 30%*. This shows that the interpretable graph representations learnt by GSAT do not generalize to larger graphs, thus undermining GSAT interpretability while *strengthening our contribution*. 2- **Novelty in method**: iGNNs differ from GSAT from a methodological perspective. 3- **Novelty in application**: to the authors knowledge, our paper is the first one to apply AI methods to investigate lattice equational properties through graphs and, most importantly, to propose a graph-based neural model capable of automatically identifying known conjectures in universal algebra. Furthermore, we remark that iGNNs are not ad-hoc models for universal algebra^, as they can be applied on top of (almost) any GNN to make it interpretable through the identification of relevant sub-patterns in a graph classification task. ^if we interpret the reviewer's concern “*GSAT is also ad-hoc mechanism*” correctly.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC