Pith. sign in

REVIEW 4 major objections 4 minor 65 references

Inference-friendly Graph Compression for Graph Neural Networks

T0 review · 4 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read This paper claims that graph compression can be inference-friendly: merging structurally indistinguishable nodes yields a compressed graph on which any matching GNN's inference returns exactly the original output.

desk verdict Clever compression idea with a load-bearing exactness proof that doesn't hold for degree- or attention-weighted GNNs; the approximate variant might be salvageable. read the letter →

arxiv 2504.13034 v3 pith:MOGYXZSK submitted 2025-04-17 cs.LG

classification cs.LG
keywords graphneuralnetworksinference-friendlycompressioninferenceequivalencestructuralcoarseningquotientmemoizationGNNacceleration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that graph compression can be made inference-friendly: instead of approximating the graph and re-running a GNN, one can merge nodes that are provably indistinguishable to a whole class of GNNs and then run inference directly on the compressed graph, recovering the original output by rescaling with cached statistics. The payoff is that one offline compression serves many inference queries and many GNNs with the same update form, and only the compressed graph is touched at query time. The paper gives three concrete schemes—SPGC with exact no-decompression inference, $(\alpha,r)$-SPGC with a tunable compression-quality trade-off, and anchored SPGC for a designated set of nodes—and reports large speedups on real and synthetic graphs. The load-bearing claim is Theorem 3: structural equivalence is an inference equivalence relation, so the quotient graph plus a scaling-factor-restoring post-processor is an IFGC.

What carries the argument

The central object is the inference equivalence relation $r_f^c$: node pairs whose embeddings coincide at every layer for every GNN in the class. The practical engine is structural equivalence $r_m$—two nodes are equivalent when features match and each neighbor of one has a matching neighbor of the other under the same relation. SPGC computes the maximum such relation with a DPP-style partition refinement, builds the quotient graph, and writes a memoization table whose entries are scaling factors (degrees, edge weights, counts) used by the rewritten update function. The theorem does the work: $r_m$ is claimed to be an inference equivalence relation, so Lemma 2 turns any quotient graph of it into an IFGC.

What would settle it

Run a fixed two-layer GCN on a graph with two nodes $u,v$ that have identical input features and bisimilar neighborhoods but different degrees (for instance, $u$ in a 2-cycle and $v$ in a 3-cycle). SPGC would merge them. Compute the exact first-layer embeddings: if $h_u^{(1)} \neq h_v^{(1)}$ for some weight matrix, then structural equivalence is not an inference equivalence relation for the GCN class, contradicting Theorem 3.

Watch

Extended reading notes

Core claim

The paper's central discovery is that inference of a whole class of GNNs can be made invariant to merging exactly the node pairs that are structurally equivalent: nodes with identical input features and neighborhoods that can be matched pairwise by the same relation. Merging those pairs yields a quotient graph $G_c$; a memoization table caches the neighbor statistics (degrees, edge weights, counts) that the update function needs, and a rewritten update function runs directly on $G_c$ and rescales embeddings to recover the original outputs. The paper claims this gives $M(G) = M(P(G_c))$ for every GNN $M$ in the class, with no decompression, and that the resulting compressed graph is the unique smallest one up to isomorphism.

Load-bearing premise

The guarantee stands on the premise that matched neighbors with equal intermediate embeddings always contribute identically to the next layer; for GCN and GAT, contributions also depend on degrees or learned attention weights, which the matching does not force to be equal.

Editorial extensions

If this is right

  • If SPGC is correct, any GNN in the class can answer queries on $G_c$ alone, with outputs restored from the memoization table, so inference cost scales with the compressed graph rather than the original.
  • The once-for-all nature means the same compressed graph serves many GNNs that share the node-update form and many query node sets, making the compression cost amortize over large workloads.
  • Anchored compression gives a targeted version: outputs for a designated anchor set are preserved with cost bounded by the $L$-hop neighborhood of those anchors, so queries on a few nodes of interest compress far more aggressively.
  • Configurable $(\alpha,r)$-compression trades exactness for size: larger similarity thresholds and wider hop ranges increase compression but require partial decompression at inference time.
  • SPGC preserves the discriminative set of the GNN class, so graph-level questions about which graphs receive the same embeddings can be answered on compressed graphs without changing the class's answer.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the exactness proof assumes matched neighbors contribute equally at the next layer, which is true for sum or average aggregation with identical coefficients but not for degree-normalized GCN or attention-based GAT; enforcing equal degrees or equal attentions inside the equivalence relation would repair the proof but shrink the compression ratio.
  • Editorial extension: the worst-case bound for $(\alpha,r)$-SPGC with decompression is the same order as inference on the original graph; the practical speedup therefore lives in how often the decompression routine can stop early, which the experiments measure but the theory does not guarantee.
  • Editorial extension: the framework suggests a simple empirical test for graph compressibility—count nodes that are role-equivalent under structural equivalence; real-world networks like citation graphs already show large fractions, so the approach transfers to settings where role structure is strong.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes inference-friendly graph compression (IFGC): given a graph G and a class of GNNs M_c, a compressor C produces a smaller graph G_c and a post-processor P such that M(G) = M(P(G_c)) for every GNN M in M_c. It introduces an inference-equivalence relation, claims that structural equivalence is inference equivalence (Theorem 3), and instantiates the framework as structural-preserving compression (SPGC), a configurable (α,r)-SPGC, and anchored SPGC (ASPGC), together with memoization tables for scaling factors. Experiments on Cora, Arxiv, Yelp, Products, and billion-scale synthetic graphs report compression ratios, inference speedups, and accuracy/F1 comparisons against DSpar and FGC. The manuscript is a combined technical report containing two full-length works: an ExGIS system paper and the IFGC paper under revision.

Significance. If the exactness theorem were sound, the paper would offer a valuable model-agnostic, once-for-all compression scheme for GNN inference, with linear-time compression and inference cost depending on the compressed graph rather than the original graph. The algorithmic ideas are clearly presented, and the large-scale experimental study is a strength. However, the central exactness guarantee is not valid as stated: structural equivalence does not preserve degree-normalized or attention-weighted aggregation for the very GNN classes in Table 3, and the key proofs are deferred to an external full version. The contribution is therefore currently unsupported, despite the attractive framing and extensive evaluation.

major comments (4)
  1. [Section 4.1, Theorem 3]
  2. [Section 3, Lemma 2 and Table 4]
  3. [Section 4.2 and Appendix A]
  4. [Section 6, Fig. 7 and Appendix C, Table 9]
minor comments (4)
  1. [Overall structure]
  2. [Section 5.1, (α,r)-relation]
  3. [Table 4 and Figure 4]
  4. [Section 2, GNN classes]

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the compression guarantees rest on structural-equivalence definitions and explicit scaling-factor rewrites, not on fitted predictions or load-bearing self-citations; the main proof gap is a correctness concern rather than circular reasoning.

full rationale

Most of the derivation chain is definitional and constructive rather than circular. The inference-equivalence relation is defined as embedding equality for all GNNs in the class, and Lemma 2's sufficient condition is a direct criterion: if C produces the quotient of such a relation and P restores embeddings via scaling factors, then M(G)=M(P(G_c)) follows by the induction sketched in the appendix; this is a sufficiency theorem, not a renamed prediction. SPGC's r_m is standard structural equivalence/bisimulation with an added feature-equality constraint, and the scaling factors in Table 4 are derived from the node-update functions (degrees for GCN, attention weights for GAT, counts for GIN), not fitted to target inference outputs. The alpha and r thresholds are user-configurable inputs, not fitted parameters renamed as predictions. The only citation-burden issue is that the paper repeatedly defers proofs and extra experiments to the authors' own full version [1]; however, proof sketches and the scaling-factor rewrites are present in the main text, so the self-citation is not load-bearing. The skeptic's concern about Theorem 3—that structural equivalence does not guarantee equal GCN/GAT aggregation coefficients—is a correctness gap in the proof sketch, which asserts aggregation equivalence from matched intermediate embeddings without requiring matched degrees or attention weights; it is not a circularity, because no equation in the paper reduces the theorem to its own inputs by construction. Accordingly the circularity score is low.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central derivation rests on structural equivalence and the scaling-factor rewrite. The main uncharged assumptions are the equality of aggregated messages for matched neighbors, availability of GAT edge weights, and near-linear DPP correctness. The paper does not pay for these assumptions independently, and the first is false for degree-normalized or attention models as stated.

free parameters (3)
  • Similarity threshold alpha in (alpha, r)-SPGC = 0.25, 0.5 used in experiments
    Configurable trade-off between compression ratio and inference quality; values are chosen by the user, not derived. Section 5.1 and Exp-2.
  • Hop radius r in (alpha, r)-SPGC = 1, 2, 3 used in experiments
    Number of hops over which neighbors are matched; larger r yields more compression and slight accuracy loss. Section 5.1 and Exp-2.
  • Feature discretization granularity = unspecified
    Appendix D says input features are discretized before similarity merging; the bin width or grouping rule is not specified and affects which nodes merge.
assumptions (5)
  • domain assumption GNNs under consideration have fixed, deterministic inference processes, and a GNN class M_c contains GNNs with the same form of node update function and only different weights.
    Section 2 defines classes; Lemma 2 and Theorem 3 rely on this to compare outputs across models in the class.
  • domain assumption Matched neighbors with equal intermediate embeddings imply equal aggregated messages for every node update function in the class.
    Theorem 3 proof sketch: 'This ensures the equivalence of aggregated embedding computed by the node update function at X and p(X).' False for degree-normalized or attention-weighted updates unless extra invariants hold.
  • domain assumption For weight-sensitive GNNs such as GAT, all edge attention weights are available and can be cached in T at compression time.
    Section 4.3 CompressG 'weight sensitive' branch; without this, exact inference accuracy is not preserved.
  • standard math DPP algorithm computes the maximum structural equivalence relation and runs in near-linear time.
    SPGC algorithm invokes DPP and reuses its correctness; complexity claims O(|V|+|E|) depend on this. Reference [16].
  • domain assumption The scaling-factor rewrite in Table 4 exactly simulates the original update function over compressed neighbors with constant-time lookups.
    Lemma 2 and Section 4.4 inference process; for GCN the per-node coefficients are not shown to be computable in O(|G_c|) time.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inference-friendly Graph Compression for Graph Neural Networks." pith.science (2026). https://pith.science/paper/MOGYXZSK

@misc{pith2026250413034,
  author       = {Pith},
  title        = {Pith review of: Inference-friendly Graph Compression for Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MOGYXZSK}},
  note         = {Machine review of arXiv:2504.13034}
}
abstract

Graph Neural Networks (GNNs) have demonstrated promising performance in graph analysis. Nevertheless, the inference process of GNNs remains costly, hindering their applications for large graphs. This paper proposes inference-friendly graph compression (IFGC), a graph compression scheme to accelerate GNNs inference. Given a graph $G$ and a GNN $M$, an IFGC computes a small compressed graph $G_c$, to best preserve the inference results of $M$ over $G$, such that the result can be directly inferred by accessing $G_c$ with no or little decompression cost. (1) We characterize IFGC with a class of inference equivalence relation. The relation captures the node pairs in $G$ that are not distinguishable for GNN inference. (2) We introduce three practical specifications of IFGC for representative GNNs: structural preserving compression (SPGC), which computes $G_c$ that can be directly processed by GNN inference without decompression; ($\alpha$, $r$)-compression, that allows for a configurable trade-off between compression ratio and inference quality, and anchored compression that preserves inference results for specific nodes of interest. For each scheme, we introduce compression and inference algorithms with guarantees of efficiency and quality of the inferred results. We conduct extensive experiments on diverse sets of large-scale graphs, which verifies the effectiveness and efficiency of our graph compression approaches.

Figures

Figures reproduced from arXiv: 2504.13034 by the authors.

Figure 1
Figure 1. GNN-based anomaly detection in a bitcoin transaction network 𝑁. (1) ExGIS conducts an “once-for-all” compression to gen￾erate and distribute a compressed graph 𝑁𝐿 (top right) for online parallel inference query workload. (2) ExGIS processes fast online inference that detects “illicit” IPs by only consulting 𝑁𝐿 , without decompression. (3) For a user designated account 𝑂𝑀 (in green box), its online, parallel explaine… view at source ↗
Figure 2
Figure 2. ExGIS Work!ow: Overview The inference process of a GNN M with 𝐿 layers takes as input a graph 𝑀 = (𝑈,𝑉), and computes the embedding 𝑌𝑆 𝑂 for each node 𝑃 → 𝑅 at each layer 𝑋 → [1, 𝐿], by recursively applying the node update function. A GNN M has a !xed inference process, if its node update function is speci"ed by "xed input model parameters, layer number, and aggregator. It has a deterministic inference process, if 𝑂… view at source ↗
Figure 3
Figure 3. Para-ExGIS: Parallel Inference and Explanation [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (20 more)
Figure 4
Figure 4. Figure 4: Parallel Inference and Explainer with a running example: 𝑗𝑉 → GCN2; 𝑐𝑊 = {𝑜1,𝑝1 }; T1, T2: partitioned subsets of T for 𝑜1 and 𝑝1; 𝑁𝑋1 , 𝑁𝑋2 : explanation graphs for 𝑜1 and 𝑝1 and induces 𝑀𝑀1 . It creates an inference joblet 𝑙1 =< T1,𝑀𝑀1 , 𝑂𝐿 > and invokes procedure Pa…
Figure 5
Figure 5. Figure 5: Para-ExGIS: Scalability and Accuracy. ExGIS consistently achieves competitive or superior accuracy. For example, on the Yelp dataset, where information loss due to com￾pression is most severe, ExGIS still retains a signi"cantly higher F1-score (e.g., 0.53 with GAT) com…
Figure 6
Figure 6. Figure 6: ExGIS Architecture them, GCN achieves the highest speed-up (up to 19.34 with 20 processors), bene"ting from its simple and layer-wise message aggregation which scales well in parallel execution. GraphSAGE follows with moderate performance due to its localized sampling …
Figure 7
Figure 7. Figure 7: Visual Interfaces of ExGIS [6] Tingyang Chen, Dazhuo Qiu, Yinghui Wu, Arijit Khan, Xiangyu Ke, and Yunjun Gao. 2024. View-based explanations for graph neural networks. Proceedings of the ACM on Management of Data 2, 1 (2024), 1–27. [7] Agostino Dovier, Carla Piazza, an…
Figure 1
Figure 1. Figure 1: Compression Scheme to scale node classi!cation. inference queries that specify any GNN M ↑ M and any 𝑅𝑀 from 𝐿, one only need to compute 𝐿𝐿 once, to reduce the total inference cost of the workload. These bene!ts applications in large-scale tests over large graphs, real…
Figure 3
Figure 3. Figure 3: Algorithm SPGC such that 𝐿 and 𝐿⇒ are isomorphic [5]. In other words, these GNNs can “solve” graph isomorphic problem: one can issue a Boolean inference query to test if an input pair of graphs are isomorphic. Given a set of graphs G, denote the set of corresponding co…
Figure 5
Figure 5. Figure 5: Run-time generation of Memoization structure [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Compressing 𝐿5 with (0.5, 2)-SPGC. query outputs. Users may also want to con!gure the compression schemes to balance among accuracy and speed up, or to contextu￾alize the compression with inference queries that speci!es a set of test nodes 𝑅𝑀 ↓ 𝑅 of interests, such tha…
Figure 7
Figure 7. Figure 7: Comparison of (0.5, 1)-SPGC with the Baselines in Inference Speed-up and Inference Accuracy/F1-score. (the inference accuracy of ASPGC marked in colored lines, e.g., the red lines show inference accuracy of ASPGC using GCN; the inference accuracy on original graph 𝐿 ma…
Figure 8
Figure 8. Figure 8: Varying Num. Layers in GNNs (Arxiv). (a) 𝑌 v.s. ncr (b) 𝑌 v.s. Speed-up (Arxiv) (c) 𝑌 v.s. F1-score (Yelp) (d) 𝑌 v.s. Accuracy (Products) (e) 𝑍 v.s. Speed-up (f) 𝑍 v.s. Accuracy [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 10
Figure 10. Figure 10: E"ciency and Scalability at Billion scale. inference speed-up achieved by (0.25,𝑂)-SPGC for all GNNs classes notably increased. Indeed, larger 𝑂 allows (𝑁,𝑂)-SPGC to !nd and merge more node pairs with equivalent embeddings, which may not be direct neighbors of another…
Figure 9
Figure 9. Figure 9: Varying 𝑁 and 𝑂 in (𝑁,𝑂)-SPGC Varying 𝑁. Fixing 𝑂 = 1, we varied 𝑁 from 0.2 to 1, and report the results in Figs. 9(a) to 9(d). It tells us the followings. (1) As 𝑁 is increased from 0.2 to 1, ncr drops as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 11
Figure 11. Figure 11: Time Cost Analysis (left: compression cost com [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 15
Figure 15. Figure 15: Algorithm decompG in dropped inference accuracy since inference equivalence is not directly preserved. A pair (𝑋,𝑋⇒ ) in an (𝑁,𝑂)-relation 𝑕(𝑌,𝑍) is no longer conform to embedding equivalence, thus an (𝑁,𝑂)-SPGC (C, _) alone is not an IFGC, i.e., no longer inference-f…
Figure 13
Figure 13. Figure 13: Procedure DPP Algorithm 4 : (𝑁,𝑂)-SPGC Input: Graph 𝐿, node feature matrix 𝑃, con!guration (xsim, 𝑁,𝑂); a class of GNNs M𝑐 with node update function 𝑀; Output: A compressed graph 𝐿𝐿 and T⇒ , 𝑆𝑜, compressed encod￾ings ALc and rules; 1: set 𝑕(𝑌,𝑍) := ↙; set 𝑆𝑜 := {𝑅 }; …
Figure 14
Figure 14. Figure 14: Algorithm (𝑁,𝑂)-SPGC there are edges incident to nodes in 𝑛↔′ and updates 𝑛𝑜 accord￾ingly (line 8-11). It next iterates over the rank equal to 0,..., 𝑞 (line 12). Within each iteration, it conducts the followings: 1) it computes the 𝑡𝑜 and re!nes it using Paige-Tarjan…
Figure 16
Figure 16. Figure 16: Illustration of the compression by Re ↔ Pair and decompression by decompG (𝐿5 in [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: Compressing graph 𝐿6 with anchored SPGC: for 2-layered GNNs, with anchored set 𝑅𝑑 = {𝑖1}. inference. This ensures that lost information in compression can be recovered by added neighbor nodes and edges (decompression). Inference Cost. decompG is in 𝑉(|𝑅 |𝑂) time. It t…
Figure 18
Figure 18. Figure 18: Varying 𝑁 in SPGC. and its compressed counterpart 𝐿𝐿 derived using SPGC. Following the de!nition of ncr, we de!ne Memory Compression Rate (mcr) as mcr = 1 ↔ |𝑓𝐿 | |𝑓 | . It quanti!es the fraction of memory cost that is “reduced”: the larger, the better; The detailed r…
Figure 3
Figure 3. Figure 3: ) Since the original graph 𝐿 may contain nodes with distinct feature vectors, we apply a featurization pre-processing step to discretize input features. This ensures that numerical values within same range are categorized together after featurization, improving consist…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 49 canonical work pages

  1. [10]

    Jeroen Bollen, Jasper Steegmans, Jan Van den Bussche, and Stijn Vansummeren

  2. [51]

    Fengli Xu, Quanming Yao, Pan Hui, and Yong Li. 2021. Automorphic equivalence- aware graph neural network. NeurIPS 34 (2021), 15138–15150

  3. [1]

    Full version

    2025. Full version. http://github.com/Yangxin666/SPGC/blob/main/SPGC_full. pdf

  4. [2]

    Arman Ahmed, Sajan K Sadanandan, Shikhar Pandey, Sagnik Basumallik, Anurag K Srivastava, and Yinghui Wu. 2022. Event Analysis in Transmission Sys- tems Using Spatial Temporal Graph Encoder Decoder (STGED).IEEE Transactions on Power Systems (2022)

  5. [3]

    Francesco Andreuzzi. 2021. BisPy: Bisimulation in Python. Journal of Open Source Software (2021)

  6. [4]

    Adnan Aziz, Vigyan Singhal, Felice Balarin, Robert K Brayton, and Alberto L Sangiovanni-Vincentelli. 1994. Equivalences for fair kripke structures. In Au- tomata, Languages and Programming: 21st International Colloquium, ICALP 94 Jerusalem, Israel, July 11–14, 1994 Proceedings 21 . 364–375

  7. [5]

    Waïss Azizian and Marc Lelarge. 2021. Expressive Power of Invariant and Equivariant Graph Neural Networks. In ICLR

  8. [6]

    Pablo Barceló, Egor V Kostylev, Mikaël Monet, Jorge Pérez, Juan L Reutter, and Juan-Pablo Silva. 2020. The expressive power of graph neural networks as a query language. SIGMOD Record (2020)

Show all 65 references
  1. [7]

    Danielle S Bassett and Edward T Bullmore. 2017. Small-world brain networks revisited. The Neuroscientist (2017)

  2. [8]

    Maciej Besta and Torsten Hoe$er. 2018. Survey and taxonomy of lossless graph compression and space-e#cient graph representations. arXiv preprint arXiv:1806.01799 (2018)

  3. [9]

    Maciej Besta and Torsten Hoe$er. 2024. Parallel and distributed graph neural networks: An in-depth concurrency analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

  4. [11]

    Stephen P Borgatti and Martin G Everett. 1992. Notions of position in social network analysis. Sociological methodology (1992)

  5. [12]

    Linfeng Cao, Haoran Deng, Yang Yang, Chunping Wang, and Lei Chen. 2024. Graph-Skeleton:˜ 1% Nodes are Su#cient to Represent Billion-Scale Graph. In Proceedings of the ACM on Web Conference 2024 . 570–581

  6. [13]

    Ming Chen, Zhewei Wei, Bolin Ding, Yaliang Li, Ye Yuan, Xiaoyong Du, and Ji-Rong Wen. 2020. Scalable graph neural networks via bidirectional propagation. NeurIPS (2020)

  7. [14]

    Tianlong Chen, Yongduo Sui, Xuxi Chen, Aston Zhang, and Zhangyang Wang

  8. [15]

    Francisco Claude and Gonzalo Navarro. 2007. A fast and compact Web graph representation. In International Symposium on String Processing and Information Retrieval. 118–129

  9. [16]

    Agostino Dovier, Carla Piazza, and Alberto Policriti. 2001. A fast bisimulation algorithm. In Proceedings of the 13th International Conference on Computer Aided Veri!cation

  10. [17]

    Pierre Dupont, Jérôme Callut, Grégoire Dooms, Jean-Noël Monette, Yves Deville, and B Sainte. 2006. Relevant subgraph extraction from random walks in a graph. Universite Catholique de Louvain, UCL/INGI, Number RR 7 (2006)

  11. [18]

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin

  12. [19]

    Yangxin Fan, Xuanji Yu, Raymond Wieser, David Meakin, Avishai Shaton, Jean- Nicolas Jaubert, Robert Flottemesch, Michael Howell, Jennifer Braid, et al . 2023. Spatio-Temporal Denoising Graph Autoencoders with Data Augmentation for Photovoltaic Data Imputation. SIGMOD (2023)

  13. [20]

    Matthias Fey and Jan E. Lenssen. 2019. Fast Graph Representation Learning with PyTorch Geometric. In ICLR Workshop on Representation Learning on Graphs and Manifolds

  14. [21]

    Xinyi Gao, Wentao Zhang, Yingxia Shao, Quoc Viet Hung Nguyen, Bin Cui, and Hongzhi Yin. 2022. E#cient Graph Neural Network Inference at Large Scale. arXiv preprint arXiv:2211.00495 (2022)

  15. [22]

    Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nel- son Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. 2018. Al- lennlp: A deep semantic natural language processing platform. arXiv preprint arXiv:1803.07640 (2018)

  16. [23]

    Floris Geerts. 2023. A Query Language Perspective on Graph Learning. In PODS

  17. [24]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. NeurIPS (2017)

  18. [25]

    Xueting Han, Zhenhuan Huang, Bang An, and Jing Bai. 2021. Adaptive transfer learning on graph neural networks. In KDD

  19. [26]

    Mohammad Hashemi, Shengbo Gong, Juntong Ni, Wenqi Fan, B Aditya Prakash, and Wei Jin. 2024. A Comprehensive Survey on Graph Reduction: Sparsi!cation, Coarsening, and Condensation. arXiv preprint arXiv:2402.03358 (2024)

  20. [27]

    Weihua Hu, Matthias Fey, Hongyu Ren, Maho Nakata, Yuxiao Dong, and Jure Leskovec. 2021. OGB-LSC: A Large-Scale Challenge for Machine Learning on Graphs. NeurIPS (2021)

  21. [28]

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open graph benchmark: Datasets for machine learning on graphs. NeurIPS (2020)

  22. [29]

    Zengfeng Huang, Shengzhong Zhang, Chong Xi, Tang Liu, and Min Zhou. 2021. Scaling up graph neural networks via graph coarsening. In SIGKDD

  23. [30]

    Ahmad Maroof Karimi, Yinghui Wu, Mehmet Koyuturk, and Roger H French

  24. [31]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classi!cation with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

  25. [32]

    Manoj Kumar, Anurag Sharma, Shashwat Saxena, and Sandeep Kumar. 2023. Featured graph coarsening with similarity guarantees. PMLR

  26. [33]

    at. 2000. O

    N Jesper Larsson and Alistair Mo"at. 2000. O"-line dictionary-based compression. Proc. IEEE 88, 11 (2000), 1722–1732

  27. [34]

    Spatiotemporal graph neural network for performance prediction of photovoltaic power systems. In AAAI

  28. [35]

    Xin Liu, Mingyu Yan, Lei Deng, Guoqi Li, Xiaochun Ye, Dongrui Fan, Shirui Pan, and Yuan Xie. 2022. Survey on graph neural network acceleration: An algorithmic perspective. arXiv preprint arXiv:2202.04822 (2022)

  29. [36]

    Yajiong Liu, Yanfeng Zhang, Qiange Wang, Hao Yuan, Xin Ai, and Ge Yu. 2025. NeutronSketch: An in-depth exploration of redundancy in large-scale graph neural network training. Knowledge-Based Systems 309 (2025), 112786

  30. [37]

    Zirui Liu, Kaixiong Zhou, Zhimeng Jiang, Li Li, Rui Chen, Soo-Hyun Choi, and Xia Hu. 2023. DSpar: An Embarrassingly Simple Strategy for E#cient GNN Training and Inference via Degree-Based Sparsi!cation. TMLR (2023)

  31. [38]

    Dongyue Li, Tao Yang, Lun Du, Zhezhi He, and Li Jiang. 2021. AdaptiveGCN: E#cient GCN through adaptively sparsifying graphs. In CIKM

  32. [39]

    Andreas Loukas and Pierre Vandergheynst. 2018. Spectrally approximating large graphs with smaller graphs. In ICML

  33. [40]

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore

  34. [41]

    Hongwu Peng, Deniz Gurevin, Shaoyi Huang, Tong Geng, Weiwen Jiang, Orner Khan, and Caiwen Ding. 2022. Towards sparsi!cation of graph neural networks. In 2022 IEEE 40th International Conference on Computer Design (ICCD) . 272–279

  35. [42]

    Francois Lorrain and Harrison C White. 1971. Structural equivalence of indi- viduals in social networks. The Journal of mathematical sociology 1, 1 (1971), 49–80

  36. [43]

    Patrick Reiser, Marlen Neubert, André Eberhard, Luca Torresi, Chen Zhou, Chen Shao, Houssam Metni, Clint van Hoesel, Henrik Schopmans, Timo Sommer, et al

  37. [44]

    Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. 2021. E (n) equi- variant graph neural networks. In ICML

  38. [45]

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. 2008. The graph neural network model. IEEE transac- tions on neural networks (2008)

  39. [46]

    Weijing Shi and Raj Rajkumar. 2020. Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud. In CVPR

  40. [47]

    Hao Peng, Hongfei Wang, Bowen Du, Md Zakirul Alam Bhuiyan, Hongyuan Ma, Jianwei Liu, Lihong Wang, Zeyu Yang, Linfeng Du, Senzhang Wang, et al

  41. [48]

    Petar Veli%kovi&. 2023. Everything is connected: Graph neural networks. Current Opinion in Structural Biology (2023)

  42. [49]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. 2017. Graph attention networks. stat (2017)

  43. [50]

    Duncan J Watts and Steven H Strogatz. 1998. Collective dynamics of ‘small- world’networks. nature (1998)

  44. [52]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2019. How powerful are graph neural networks?. In ICLR

  45. [53]

    Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, and Ren Chen. 2021. Decoupling the depth and scope of graph neural networks. NeurIPS (2021)

  46. [54]

    Shyam A Tailor, Javier Fernandez-Marques, and Nicholas D Lane. 2020. Degree- quant: Quantization-aware training for graph neural networks. arXiv preprint arXiv:2008.05000 (2020)

  47. [55]

    Muhan Zhang and Yixin Chen. 2018. Link prediction based on graph neural networks. NeurIPS (2018)

  48. [56]

    Cheng Zheng, Bo Zong, Wei Cheng, Dongjin Song, Jingchao Ni, Wenchao Yu, Haifeng Chen, and Wei Wang. 2020. Robust graph representation learning via neural sparsi!cation. In ICML

  49. [57]

    W ANG Zhili, DI Shimin, CHEN Lei, and ZHOU Xiaofang. 2024. Search to !ne- tune pre-trained graph neural networks for graph-level tasks. In ICDE. Yangxin Fan, Haolai Che, Yinghui Wu

  50. [58]

    matching relation

    Hongkuan Zhou, Ajitesh Srivastava, Hanqing Zeng, Rajgopal Kannan, and Viktor Prasanna. 2021. Accelerating large scale real-time GNN inference using channel pruning. arXiv preprint arXiv:2105.04528 (2021). 8 APPENDIX 8.1 Appendix A: Proofs Proof of Lemma 1 . Given M and 𝐿, the ...

  51. [61]

    Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. 2019. Graphsaint: Graph sampling based inductive learning method. arXiv preprint arXiv:1907.04931 (2019)

  52. [2000]

    Information Retrieval (2000)

    Automating the construction of internet portals with machine learning. Information Retrieval (2000)

  53. [2019]

    Graph neural networks for social recommendation. In WWW

  54. [2020]

    Information Sciences (2020)

    Spatial temporal incidence dynamic graph neural networks for tra#c $ow forecasting. Information Sciences (2020)

  55. [2021]

    A uni!ed lottery ticket hypothesis for graph neural networks. In ICML

  56. [2022]

    Communica- tions Materials (2022)

    Graph neural networks for materials science and chemistry. Communica- tions Materials (2022)

  57. [2023]

    In Proceedings of the 6th Joint Workshop on Graph Data Management Experiences & Systems (GRADES) and Network Data Analytics (NDA)

    Learning graph neural networks using exact compression. In Proceedings of the 6th Joint Workshop on Graph Data Management Experiences & Systems (GRADES) and Network Data Analytics (NDA) . 1–9

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.