Abstract
Graph neural networks (GNNs) have shown their superiority in the modeling of graph data. Recently, increasing attention has been paid to automatic graph neural architecture search, aiming to overcome the shortcomings of manually constructing GNN architectures that requires a lot of expert experience. However, existing graph neural architecture search (GraphNAS) methods can only select architecture from the partial evaluated GNN architectures. To solve the challenges, we propose a Graph Neural Architecture Prediction (GraphNAP) framework, which can select the optimal GNN architecture from the search space efficiently. To achieve this goal, a neural predictor is designed in GraphNAP. Firstly, the neural predictor is trained by a small number of sampled GNN architectures. Then, the trained neural predictor is used to predict all GNN architectures in the search space. In this way, GraphNAP can efficiently explore the performance of all GNN architectures in the search space and then select the optimal GNN architecture. The experimental results show that GraphNAP outperforms state-of-the-art both handcrafted and GraphNAS-based methods for both graph and node classification tasks. The python implementation of GraphNAP can be found at https://github.com/BeObm/GraphNAP.
Similar content being viewed by others
Data availability
All datasets used in the paper are publicly available.
References
Wu Z, Pan S, Chen F, Long G, Zhang C, Philip SY (2020) A comprehensive survey on graph neural networks. IEEE Trans Neural Netw Learn Syst 32(1):4–24
Bai L, Cui L, Jiao Y, Rossi L, Hancock ER (2022) Learning backtrackless aligned-spatial graph convolutional networks for graph classification. IEEE Trans Pattern Anal Mach Intell 44(2):783–798
Manipur I, Manzo M, Granata I, Giordano M, Maddalena L, Guarracino MR (2021) Netpro2vec: a graph embedding framework for biomedical applications. IEEE/ACM Trans Comput Biol Bioinf 19(2):729–740
Gao Y, Yang H, Zhang P, Zhou C, Hu Y (2020) Graph neural architecture search. In: Proceedings of the joint conference on artificial intelligence, pp 1403–1409
Pasa L, Navarin N, Sperduti A (2022) Polynomial-based graph convolutional neural networks for graph classification. Mach Learn 111:1205–1237
Chen J, Gao J, Chen Y, Moctard OB, Lyu T, Li Z (2022) Auto-GNAS: a parallel graph neural architecture search framework. IEEE Trans Parallel Distrib Syst 33(11):3117–3128
Qin Y, Wang X, Zhang Z, Zhu W (2021) Graph differentiable architecture search with structure learning. In: Proceedings of the annual conference on neural information processing systems, pp 16860–16872
Guan C, Wang X, Zhu W (2021) Autoattend: automated attention representation search. In: Proceedings of the international conference on machine learning, pp 3864–3874
Li Y, King I (2020) Autograph: automated graph neural network. In: Proceedings of the international conference on neural information processing. Springer, pp 189–201
Guo Z, Zhang X, Mu H, Heng W, Liu Z, Wei Y, Sun J (2020) Single path one-shot neural architecture search with uniform sampling. In: Proceedings of the European conference on computer vision. Lecture notes in computer science, pp 544–560
Li Y, Wen Z, Wang Y, Xu C (2021) One-shot graph neural architecture search with dynamic search space. In: Proceedings of the AAAI conference on artificial intelligence, pp 8510–8517
Gao Y, Yang H, Zhang P, Zhou C, Hu Y (2020) Graph neural architecture search. In: Proceedings of the international joint conference on artificial intelligence, pp 1403–1409
Zhou K, Song Q, Huang X, Hu X (2019) Auto-gnn: neural architecture search of graph neural networks. CoRR arXiv:1909.03184
You J, Ying Z, Leskovec J (2020) Design space for graph neural networks. In: Proceedings of advances in neural information processing systems, pp 17009–17021
Sperduti A, Starita A (1997) Supervised neural networks for the classification of structures. IEEE Trans Neural Netw 8(3):714–735
Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl GE (2017) Neural message passing for quantum chemistry. In: Proceedings of the international conference on machine learning. PMLR, pp 1263–1272
Kolouri S, Naderializadeh N, Rohde GK, Hoffmann H (2021) Wasserstein embedding for graph learning. In: Proceedings of the international conference on learning representations
Kipf TN, Welling M (2017) Semi-supervised classification with graph convolutional networks. In: Proceedings of the international conference on learning representations
Wang Y, Wang W, Liang Y, Cai Y, Hooi B (2021) Curgraph: curriculum learning for graph classification. In: Proceedings of the web conference, pp 1238–1248
Xu K, Hu W, Leskovec J, Jegelka S (2019) How powerful are graph neural networks? In: Proceedings of the international conference on learning representations
Oloulade BM, Gao J, Chen J, Lyu T, Al-Sabri R (2021) Graph neural architecture search: a survey. Tsinghua Sci Technol 27(4):692–708
Zhou K, Song Q, Huang X, Hu X (2019) Auto-gnn: neural architecture search of graph neural networks. CoRR arXiv:1909.03184
Sutton RS, McAllester DA, Singh SP, Mansour Y (1999) Policy gradient methods for reinforcement learning with function approximation. In: Proceedings of advances in neural information processing systems, pp 1057–1063
Yu K, Sciuto C, Jaggi M, Musat C, Salzmann M (2020) Evaluating the search phase of neural architecture search. In: Proceedings of the international conference on learning representations
Zhang Y, Lin Z, Jiang J, Zhang Q, Wang Y, Xue H, Zhang C, Yang Y (2020) Deeper insights into weight sharing in neural architecture search. CoRR arXiv:2001.01431
Shi M, Wilson DA, Zhu X, Huang Y, Zhuang Y, Liu J, Tang Y (2020) Evolutionary architecture search for graph neural networks. CoRR arXiv:2009.10199
Li Y, King I (2020) Autograph: automated graph neural network. In: Proceedings of the international conference on neural information processing, pp 189–201
Zhao H, Yao Q, Tu W (2021) Search to aggregate neighborhood for graph neural network. In: Proceedings of the IEEE international conference on data engineering, pp 552–563
Lai K-H, Zha D, Zhou K, Hu X (2020) Policy-gnn: Aggregation optimization for graph neural networks. In: Proceedings of the ACM SIGKDD international conference on knowledge discovery and data mining, pp 461–471
Shi H, Pi R, Xu H, Li Z, Kwok JT, Zhang T (2020) Bridging the gap between sample-based and one-shot neural architecture search with BONAS. In: Proceedings of annual conference on neural information processing systems, pp 6–12
White C, Zela A, Ru R, Liu Y, Hutter F (2021) How powerful are performance predictors in neural architecture search? In: Proceedings of annual conference on neural information processing systems, pp 28454–28469
White C, Neiswanger W, Savani Y (2021) BANANAS: Bayesian optimization with neural architectures for neural architecture search. In: Proceedings of AAAI conference on artificial intelligence, pp 10293–10301
Hu W, Fey M, Zitnik M, Dong Y, Ren H, Liu B, Catasta M, Leskovec J (2020) Open graph benchmark: datasets for machine learning on graphs. In: Proceedings of advances in neural information processing systems
Wei L, Zhao H, Yao Q, He Z (2021) Pooling architecture search for graph classification. In: Proceedings of the ACM international conference on information and knowledge management, pp 2091–2100
Velickovic P, Cucurull G, Casanova A, Romero A, Liò P, Bengio Y (2018) Graph attention networks. In: Proceedings of the international conference on learning representations
Huang W, Zhang T, Rong Y, Huang J (2018) Adaptive sampling towards fast graph representation learning. In: Proceedings of advances in neural information processing systems, pp 4563–4572
Wang Y, Derr T (2021) Tree decomposed graph neural network. In: Demartini G, Zuccon G, Culpepper JS, Huang Z, Tong H (eds) Proceedings of the ACM international conference on information and knowledge management, pp 2040–2049
Chen M, Wei Z, Huang Z, Ding B, Li Y (2020) Simple and deep graph convolutional networks. In: Proceedings of the international conference on machine learning, pp 1725–1735
Fey M, Lenssen JE (2019) Fast graph representation learning with PyTorch geometric. In: Proceedings of THE ICLR workshop on representation learning on graphs and manifolds
Acknowledgements
The work was supported by the National Natural Science Foundation of China (No. 62272487).
Funding
The work was supported by the National Natural Science Foundation of China (No. 62272487).
Author information
Authors and Affiliations
Contributions
J.G. did conceptualization, investigation, project administration, supervision, editing, validation, and review. B.M.O. done conceptualization, data duration, formal analysis, investigation, methodology, resources, software, validation, visualization, editing, and review. R.A. performed investigation, resources, editing, validation, and review. J.C. contributed to investigation, resources, editing, validation, and review. T.L. and Z.W. were involved in investigation, validation, and review.
Corresponding author
Ethics declarations
Conflict of interest
No conflicts of interest or competing interests.
Ethics approval
Not applicable.
Consent to participate
The manuscript is approved by all authors for publication.
Consent for publication
All authors who participated in this study give the publisher permission to publish this work.
Code availability
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
We aim to demonstrate that the CHRS method yields a more representative and diverse set of sampled architectures compared to random and uniform sampling.
1.1 Proof of diversity
We first prove that CHRS generates a more diverse set of architectures compared to uniform sampling. In uniform sampling, each candidate is equally likely to be selected, resulting in a higher chance of repeated architectures. In CHRS, we sample m GNN architectures for each candidate, ensuring a more comprehensive exploration of the search space.
Lemma 1
The CHRS method produces a set T of n GNN architectures with a higher diversity compared to uniform sampling.
Proof
Assume we have k unique architectures in the search space. In uniform sampling, the probability of selecting a specific architecture is 1/k. The probability of selecting m architectures of the same candidate is \((1/k)^m\). Thus, the probability of not selecting any repeated architectures for a candidate is given by \((1 - (1/k)^m)\). For CHRS, the lower limit m is determined as m = round(n/q). Therefore, the probability of not selecting any repeated architectures for a candidate in CHRS is \((1 - (1/k)^m)\), which is greater than or equal to the probability for uniform sampling. Since CHRS ensures a higher probability of not selecting repeated architectures, it generates a more diverse set of architectures compared to uniform sampling. Hence, Lemma 1 is proven.
1.2 Proof of representativeness
Next, we prove that CHRS provides a more representative sampling of the search space than random sampling. Random sampling lacks control over the distribution of sampled architectures, potentially leading to an insufficient representation of the search space. \(\square \)
Lemma 2
The CHRS method yields a set T of n GNN architectures that better represents the characteristics of the search space compared to random sampling.
Proof
In random sampling, each architecture has an equal chance of being selected. The probability of selecting m architectures with the same candidate is \((1/q)^m\). The probability of not selecting any repeated architectures for a candidate is \((1 - (1/q)^m)\).
In CHRS, we sample m architectures for each candidate, ensuring a controlled distribution of architectures. By sampling m architectures, we increase the chance of covering various regions of the search space for each candidate. Therefore, the probability of not selecting any repeated architectures for a candidate in CHRS is \((1 - (1/q)^m)\), which is greater than or equal to the probability for random sampling. As CHRS provides a higher probability of not selecting repeated architectures, it leads to a more representative sampling of the search space compared to random sampling. Hence, Lemma 2 is proven.
Based on Lemmas 1 and 2, we conclude that the CHRS method produces a more diverse and representative set of GNN architectures compared to random and uniform sampling methods. These properties enable CHRS to capture a broader range of architectures, ultimately leading to improved performance of the machine learning model. \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Gao, J., Oloulade, B.M., Al-Sabri, R. et al. Graph neural architecture prediction. Knowl Inf Syst 66, 29–58 (2024). https://doi.org/10.1007/s10115-023-01968-6
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10115-023-01968-6