Pith. sign in

REVIEW 4 major objections 6 minor 56 references

Improving LLM Reasoning with Homophily-aware Structural and Semantic Text-Attributed Graph Compression

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read Compressing graphs by homophily makes LLM reasoning more accurate

desk verdict A plausible homophily-based graph compression pipeline for LLM node classification whose main claim about shrinking LLM inputs is not actually measured — the reported compression is graph-node count, not prompt tokens. read the letter →

arxiv 2601.08187 v3 pith:K5USJCWU submitted 2026-01-13 cs.AI

classification cs.AI
keywords text-attributedgraphsgraphcompressionstructuralentropyhomophilyLLMreasoningnodeclassificationcommunitydetectionprompt
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

Large language models reading graph-structured text must fit the graph into a fixed context window, and existing pipelines do this by randomly dropping neighbors—a step this paper identifies as a major source of noise and unstable predictions. The paper's claim is that a graph can be compressed before the prompt is built, and that the right compression—one that detects communities of similar nodes and distills each community's background text into a short consensus—both shrinks the input and improves the answer. The proposed HS2C pipeline adds text-similarity edges, partitions the graph by minimizing structural entropy, and then uses the LLM itself to summarize each community's background nodes conditionally on the target node. Across 10 node-level and 7 graph-level benchmarks, the compressed graphs yield higher zero-shot accuracy than random, degree-based, retrieval, or skeleton-based neighbor sampling, while often reducing graph size to about 5% of the original. If the claim holds, context-window limits stop being a reason for noisy sampling: compression becomes a way to feed the LLM structure rather than discard it.

What carries the argument

The load-bearing object is the structural entropy coding tree T built by greedy MERGE and DROP operations that minimize H_T(G) = -Σ_{α≠λ} (g_α/vol(G)) log₂(vol(α)/vol(α⁻)). Structural entropy measures the uncertainty of a biased random walk over the graph; minimizing it yields a hierarchical partition whose finest level above the leaves is taken as the set of homophilic communities. This partitioning runs on a KNN-enhanced graph: BERT embeddings of node texts, Pearson-correlation similarity, and a k chosen by the plateau of one-dimensional structural entropy. The second half of the machinery is Community-variant Semantics Aggregation: each community is typed (Pure Target, Specific Target, Co

What would settle it

Construct or select a graph where node texts are deliberately similar across classes (shared keywords, different labels) and run HS2C with increasing KNN k; if accuracy falls below random-sampling baselines as k grows, the homophily proxy fails. Alternatively, measure the label purity of HS2C's detected communities on a heterophilic benchmark and check whether it still exceeds Louvain/Leiden; if it does not, the entropy-guided partition is not finding label-relevant structure.

Watch

Extended reading notes

Core claim

In the paper's own terms, HS2C is a homophily-aware framework for community-level compression of text-attributed graphs. It argues that graphs encode both structural and semantic homophily—nodes that connect tend to share topical content—and that uncovering this homophily is the key to compression without information loss. Structurally, HS2C adds k-nearest-neighbor edges based on BERT text similarity, then minimizes structural entropy to build a hierarchical coding tree; the finest partition of that tree defines the homophilic communities, and the paper shows these communities have markedly higher label purity than Louvain or Leiden partitions. Semantically, it types each community as Pure,

Load-bearing premise

The load-bearing premise is that BERT-based semantic similarity between node texts tracks label-relevant homophily, so the added KNN edges steer the structural-entropy partition toward communities whose members share the classification-relevant signal; if textual similarity and class labels are orthogonal, compression will route the wrong context into the prompt and hurt accuracy.

Editorial extensions

If this is right

  • Frozen-LLM prompting on graphs no longer has to choose between context-window fit and information; the compression step is where structure is preserved, so the prompt itself only sees a distilled graph.
  • The same compressed graph can be served to different LLM families and sizes (3B, 7B, 8B tested), since summarization happens once offline and inference reuses it.
  • Because GCR and ACC improve together, the method's practical value is measurable as accuracy per unit of input (GCI), not just as a compression trick.
  • The pipeline scales to graphs with roughly 2M nodes and 30M edges (DBLP), suggesting the approach is usable outside toy benchmarks.
  • Graph-level tasks—sentiment graphs and molecular property graphs—respond to the same community-summary recipe, so the mechanism is not node-classification-specific.

Reading between the lines

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

  • The method's boundary condition is the homophily proxy itself: if BERT text similarity is orthogonal to the label signal (e.g., papers that share vocabulary but belong to different subfields), the KNN-enhancement step can inject misleading edges, and the entropy partition would compress the wrong structure. This is a testable failure mode the paper does not explore.
  • Because community summaries are generated once with a frozen LLM, HS2C shifts compute from per-target prompting to one-time compression; whether this amortizes at industrial scale—where the graph changes over time—is an economic question the paper leaves open.
  • The four-type community taxonomy invites a refinement the paper does not pursue: weighting aggregated summaries by each community's label purity (a quantity the paper already measures in its homophily-score table) could sharpen the signal for Mixed-Target communities.
  • The paper's stated future direction of temporal graphs suggests a concrete stress test: when communities drift, a static compressed graph may encode stale homophily, so the method's stability over time is an open empirical question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes HS2C, a compression framework for text-attributed graphs used in zero-shot LLM node classification. HS2C first enhances the graph with a BERT-similarity KNN graph, then performs structural-entropy minimization to obtain a hierarchical community partition. Communities are classified into four types according to their target/background composition, and the background texts in each community are summarized by an LLM. The compressed graph retains all target nodes plus one aggregated background node per community; the LLM is then prompted with the target text and the texts of all 1-hop neighbors in this compressed graph. The authors claim that HS2C simultaneously improves compression rate and downstream accuracy, reporting consistent gains over GNN, random/degree/number/RAG sampling, graph-skeleton, and graph-coarsening baselines on 10 node-level and 7 graph-level benchmarks, with a 94.98% graph-scale compression and 3.06%–4.92% accuracy improvement on OGBN-ArXiv.

Significance. If the claims are substantiated, the paper would make a useful contribution: it provides a principled, entropy-guided way to compress TAGs for LLM prompting while claiming to improve accuracy. The strengths are the broad experimental coverage, the detailed prompt templates in the appendix, and the attempt to ground community detection in structural information theory. The central hypothesis—that homophily-aware community summaries give the LLM more relevant context than random neighbors—is plausible and worth testing. However, the current evidence does not yet establish the headline claim of 'input compression': the reported compression rate counts graph nodes, not the tokens actually fed to the LLM, and the baselines are token-budget-limited while HS2C is not. The paper also lacks significance-test details despite claiming statistical significance, has no quantitative ablation, and provides no code. These gaps are fixable within the manuscript's scope, so the appropriate decision is major revision rather than rejection.

major comments (4)
  1. [§5.1, Eq. (5), §9.2] The central claim of 'input compression' is not supported by the reported metric. GCR is defined as |eV|/|V|, i.e., node-count reduction, but the LLM input in Eq. (5) consists of the target text plus the texts of all 1-hop neighbors in the compressed graph eG. Since all target nodes are preserved and edges are redirected according to Eq. (4), a target's neighbor set in eG can contain many original target nodes and multiple parallel edges to the same supernode; the prompt token count is never reported. Meanwhile, baselines are explicitly token-budget-limited ('The overall sampling process in all methods is limited by the predefined token number of the LLM context constraint,' §9.2), while no such limit is stated for HS2C. Thus the 94.98% 'graph scale compression' on OGBN-ArXiv is not evidence that the LLM input is smaller, and the accuracy gains could partly come from giving HS2C more con
  2. [Table 1 caption; Table 2; Table 5] The Table 1 caption states that 'All performance improvements achieved by HS2C are statistically significant,' but no significance test, number of trials beyond five seeds, or test procedure is described. Many entries report standard deviations of exactly 0.00 (e.g., HS2C 3B on OGBN-ArXiv 35.01±0.00 in Table 2; several HS2C rows in Table 5), which suggests either deterministic runs or rounding that makes the significance claim unverifiable. Provide the actual per-seed results or significance tests, and soften the claim unless the tests are reported.
  3. [§5.4, Figure 5] The ablation study is essential for supporting the claim that GSE, HCP, and CSA each contribute to the final performance, but Figure 5 appears to be a qualitative illustration: no numeric accuracy values, error bars, or axis values are given. The text says removing GSE leads to an 'accuracy drop' and that components are 'mutually reinforcing,' yet the reader cannot quantify these drops or compare them across variants. Please report the ablation numbers in a table or with numerical labels in the figure.
  4. [§4.1.1, Table 4] The KNN enhancement in Eq. (2) is load-bearing: it modifies the graph on which all structural-entropy partitioning and community detection are performed. The paper selects k via an entropy criterion, but gives no sensitivity analysis with respect to k or to the choice of BERT encoder/pooling. Table 4 evaluates homophily against ground-truth labels, but the partitioning itself never uses labels; the reader cannot tell whether the reported gains are stable or dependent on a particular k/encoder setting. A sensitivity study over k and at least one alternative text encoder would substantially strengthen the claim that the detected communities are robust.
minor comments (6)
  1. [§5.1, Appendix 9.1] The paper says it evaluates on 10 node-level benchmarks, but §5.1 lists only six datasets (OGBN-ArXiv, TAPE, Instagram, Citeseer, Reddit, Product-subset) and then mentions two additional large-scale datasets in §5.5. Please make the count consistent and clarify that the large-scale results are included in the 10.
  2. [§2.2, §6, §9.2] There are several typos: 'A another' (§2.2), 'extentions' (§6), 'is is limited' (§9.2), and 'homophilc structures' (Table 4 caption). Please proofread.
  3. [§4.1.1, Eq. (2)] The BERT embedding x_i is described only as 'obtained by feeding its text attribute... into a pretrained BERT model.' For reproducibility, specify whether the [CLS] token, mean-pooled token embeddings, or another pooling is used, and which BERT variant is used.
  4. [§4.4, Eq. (4)] Because Eq. (4) redirects every edge incident to each background node in a community to the same supernode, a target can have multiple parallel edges to the same supernode. It should be clarified whether parallel edges are deduplicated before forming the neighbor set in Eq. (5); otherwise the same summary may be repeated in the prompt, inflating the token count.
  5. [§5.4, Figure 5] Even if the ablation is intended to be qualitative, the figure is currently difficult to interpret because the axes are unlabeled and no numeric values are shown. Add at least axis labels and numerical markers, or move the ablation results to a table.
  6. [General] No code is released. Given the complexity of the pipeline (KNN construction, structural-entropy partitioning, LLM summarization, compressed-graph reconstruction), releasing code and the exact prompt templates would greatly aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is empirical, and the only self-citation is non-load-bearing.

full rationale

The paper's claimed derivation chain is an empirical compression pipeline, not a formal derivation that reduces to its inputs. GSE adds KNN edges based on BERT text similarity; HCP minimizes structural entropy using the external theory of Li and Pan [13]; CSA summarizes background texts conditioned on target texts; and the compressed graph is reconstructed by redirecting edges. No parameter is fitted to the test labels, and accuracy is measured on held-out nodes. The compression metric GCR is explicitly defined as |eV|/|V|, and the reported 'compression rate' is exactly that node-count ratio, so the claim is internally consistent with its stated metric. The only self-citation, [14], appears in a list of structural-entropy applications and is not load-bearing; no uniqueness theorem is invoked, and no ansatz is smuggled via self-citation. The homophily evaluation uses labels only as an external evaluation metric, not as a fitting signal. A possible concern that GCR measures node count rather than actual prompt token count is a validity/correctness issue about whether the metric supports the abstract's 'input compression' claim, not a circularity of the derivation. Therefore no circular step is present.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

All free parameters are structural hyperparameters (k, h) chosen by unpublished heuristics. The central axioms are domain assumptions about text-similarity-to-label alignment and SE revealing essential topology; neither is proven for these datasets. The method introduces no external entity with independent evidence.

free parameters (2)
  • k (KNN neighbor count) = not reported
    Chosen per graph by the marginal-increase plateau of one-dimensional structural entropy (Sec 4.1.1); no reported values or sensitivity analysis.
  • tree height h (HCP) = not reported
    Predefined coding-tree height in Algorithm 1; no criterion or reported values, though it determines community granularity and therefore downstream compression quality.
assumptions (5)
  • domain assumption Objects with similar attributes tend to form connections (homophily assumption).
    Invoked in Sec 4.1.1 to justify KNN graph enhancement; on heterophilic graphs the added edges may be wrong and degrade the partition.
  • domain assumption Minimizing structural entropy isolates essential topology and discards stochastic noise.
    Adopted from Li & Pan [13]; used as the foundation of the hierarchical community partition in Sec 4.1.2.
  • domain assumption BERT embeddings capture label-relevant semantic similarity.
    Eq. 2 uses f_BERT(R_i) to compute Pearson similarity; no validation of embedding quality for each dataset is given.
  • domain assumption Frozen LLM summaries preserve information needed for classification.
    Sec 4.2 relies on LLM summarization of background texts; if summaries drop discriminative details, downstream accuracy suffers.
  • domain assumption The finest-granularity communities from the coding tree are natural homophilic structures.
    Sec 4.1.2 defines Chomo as the immediate partition above leaf communities; this granularity depends on tree height and greedy merge order.
invented entities (1)
  • Condensed background supernode e_b^(i)
    purpose: Represents all background nodes of a community as one summary node so the LLM prompt fits the context window.
    Internal construct of the method; it makes no external falsifiable prediction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving LLM Reasoning with Homophily-aware Structural and Semantic Text-Attributed Graph Compression." pith.science (2026). https://pith.science/paper/K5USJCWU

@misc{pith2026260108187,
  author       = {Pith},
  title        = {Pith review of: Improving LLM Reasoning with Homophily-aware Structural and Semantic Text-Attributed Graph Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K5USJCWU}},
  note         = {Machine review of arXiv:2601.08187}
}
read the original abstract

Large language models (LLMs) have demonstrated promising capabilities in Text-Attributed Graph (TAG) understanding. Recent studies typically focus on verbalizing the graph structures via handcrafted prompts, feeding the target node and its neighborhood context into LLMs. However, constrained by the context window, existing methods mainly resort to random sampling, often implemented via dropping node/edge randomly, which inevitably introduces noise and cause reasoning instability. We argue that graphs inherently contain rich structural and semantic information, and that their effective exploitation can unlock potential gains in LLMs reasoning performance. To this end, we propose Homophily-aware Structural and Semantic Compression for LLMs (HS2C), a framework centered on exploiting graph homophily. Structurally, guided by the principle of Structural Entropy minimization, we perform a global hierarchical partition that decodes the graph's essential topology. This partition identifies naturally cohesive, homophilic communities, while discarding stochastic connectivity noise. Semantically, we deliver the detected structural homophily to the LLM, empowering it to perform differentiated semantic aggregation based on predefined community type. This process compresses redundant background contexts into concise community-level consensus, selectively preserving semantically homophilic information aligned with the target nodes. Extensive experiments on 10 node-level benchmarks across LLMs of varying sizes and families demonstrate that, by feeding LLMs with structurally and semantically compressed inputs, HS2C simultaneously enhances the compression rate and downstream inference accuracy, validating its superiority and scalability. Extensions to 7 diverse graph-level benchmarks further consolidate HS2C's task generalizability.

Figures

Figures reproduced from arXiv: 2601.08187 by the authors.

Figure 1
Figure 1. The challenge and our solution. (a) Illustration of LLM reasoning instability induced by random sampling strategy. (b) Illustration of HS2C compressing the graph to generate compact LLM inputs. 1 Introduction Large language models (LLMs) have recently emerged as powerful tools for text-attributed graph (TAG) understanding, delivering strong inference performance on downstream tasks [1, 2, 3]. Existing works primaril… view at source ↗
Figure 2
Figure 2. The overall framework of the proposed HS2C, which consists of 3 modules. Firstly, we detect homophilic structures by enhancing the graph topology and minimizing SE to obtain a hierarchical community partition. Secondly, we aggregate the textual attributes of background nodes within each community to generate concise, semantically aligned summaries. Finally, we reconstruct a compressed graph Ge that preserves essenti… view at source ↗
Figure 3
Figure 3. Comparison of Graph Compression Rate (GCR), classification accuracy (ACC), and data memory (MB). [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Node-level Case study on OGBN-ArXiv dataset comparing the neighborhood of a target node before and after applying HS2C. The left side shows the dense neighbors of target node in the original graph G. Right: the compressed neighbors after homophilic structure detection …
Figure 5
Figure 5. Figure 5: Ablation study results for HS2C’s 3 key components. We observe that removing GSE leads to a accuracy drop, confirming that structure enhancement provides richer connectivity cues for subsequent partitioning. Comparing w/o GSE and w/o HCP further highlights that structu…
Figure 6
Figure 6. Figure 6: Comparison of Graph Compression Rate (GCR) on very large graph, classification accuracy (ACC), and data memory [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Graph-level task performance Comparison. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: The initial coding tree T (1) constructed from the graph G. Each node vi ∈ V is treated as an individual community ci, and all such communities are organized as leaf nodes directly connected to the root community cλ, forming the 1-level community set C (1) . 8.2 The Al…
Figure 9
Figure 9. Figure 9: Illustration of the MERGE operation. In each iteration, two child communities of the root are greedily selected and merged to achieve the maximal reduction in SE, progressively forming a binary coding tree without height limitation [PITH_FULL_IMAGE:figures/full_fig_p0…
Figure 10
Figure 10. Figure 10: Illustration of the DROP operation. A selected intermediate community node cn is removed, and all of its child communities are directly reattached to its parent c − n , thereby compressing the coding tree to satisfy the predefined height constraint h. When two communi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 2 linked inside Pith

  1. [1]

    A survey of large language models for graphs

    Xubin Ren, Jiabin Tang, Dawei Yin, Nitesh Chawla, and Chao Huang. A survey of large language models for graphs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD-24, page 6616–6626, 2024

  2. [2]

    Large language models on graphs: A comprehensive survey.IEEE Transactions on Knowledge and Data Engineering, 36(12):8622–8642, 2024

    Bowen Jin, Gang Liu, Chi Han, Meng Jiang, Heng Ji, and Jiawei Han. Large language models on graphs: A comprehensive survey.IEEE Transactions on Knowledge and Data Engineering, 36(12):8622–8642, 2024

  3. [3]

    Graphgpt: Graph instruction tuning for large language models

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Lixin Su, Suqi Cheng, Dawei Yin, and Chao Huang. Graphgpt: Graph instruction tuning for large language models. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR-24, pages 491–500, 2024

  4. [4]

    Language is all a graph needs

    Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. Language is all a graph needs. In Findings of the Association for Computational Linguistics, EACL-24, pages 1955–1973, 2024

  5. [5]

    When do LLMs help with node classification? a comprehensive analysis

    Xixi Wu, Yifei Shen, Fangzhou Ge, Caihua Shan, Yizhu Jiao, Xiangguo Sun, and Hong Cheng. When do LLMs help with node classification? a comprehensive analysis. InForty-second International Conference on Machine Learning, ICML-25, 2025

  6. [6]

    Llaga: Large language and graph assistant

    Runjin Chen, Tong Zhao, AJAY KUMAR JAISW AL, Neil Shah, and Zhangyang Wang. Llaga: Large language and graph assistant. InForty-first International Conference on Machine Learning, ICML-24, pages 7809–7823, 2024

  7. [7]

    Beyond the limits: a survey of techniques to extend the context length in large language models

    Xindi Wang, Mahsa Salmani, Parsa Omidi, Xiangyu Ren, Mehdi Rezagholizadeh, and Armaghan Eshaghi. Beyond the limits: a survey of techniques to extend the context length in large language models. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 8299–8307, 2024

  8. [8]

    Are large language models in-context graph learners?arXiv preprint arXiv:2502.13562, 2025

    Jintang Li et al. Are large language models in-context graph learners?arXiv preprint arXiv:2502.13562, 2025

Show all 56 references
  1. [9]

    Exploring the potential of large language models (llms)in learning on graphs

    Zhikai Chen, Haitao Mao, Hang Li, Wei Jin, Hongzhi Wen, Xiaochi Wei, Shuaiqiang Wang, Dawei Yin, Wenqi Fan, Hui Liu, and Jiliang Tang. Exploring the potential of large language models (llms)in learning on graphs. ACM SIGKDD Explorations Newsletter, 25(2):42–61, 2024

  2. [10]

    Talk like a graph: Encoding graphs for large language models

    Bahare Fatemi, Jonathan Halcrow, and Bryan Perozzi. Talk like a graph: Encoding graphs for large language models. InThe Twelfth International Conference on Learning Representations, ICLR-24, 2024

  3. [11]

    Graph condensation for graph neural networks

    Wei Jin, Lingxiao Zhao, Shichang Zhang, Yozen Liu, Jiliang Tang, and Neil Shah. Graph condensation for graph neural networks. InInternational Conference on Learning Representations, ICLR-22, 2022

  4. [12]

    Condensing graphs via one-step gradient matching

    Wei Jin, Xianfeng Tang, Haoming Jiang, Zheng Li, Danqing Zhang, Jiliang Tang, and Bing Yin. Condensing graphs via one-step gradient matching. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD-22, page 720–730, 2022

  5. [13]

    Structural information and dynamical complexity of networks.IEEE Transactions on Information Theory, 62(6):3290–3339, 2016

    Angsheng Li and Yicheng Pan. Structural information and dynamical complexity of networks.IEEE Transactions on Information Theory, 62(6):3290–3339, 2016

  6. [14]

    Graph out-of-distribution generalization based on structural-entropy-guided information bottleneck.ACM Transactions on Knowledge Discovery from Data, 20(1):1–34, 2025

    Zijun Di, Peng Zheng, Bin Lu, Kai Guan, Luoyi Fu, Ningdi Jin, Ye Chen, Xiaoying Gan, Lei Zhou, Xinbing Wang, and Chenghu Zhou. Graph out-of-distribution generalization based on structural-entropy-guided information bottleneck.ACM Transactions on Knowledge Discovery from Data, ...

  7. [15]

    Structural entropy guided unsupervised graph out-of-distribution detection

    Yue Hou, He Zhu, Ruomei Liu, Yingke Su, Jinxiang Xia, Junran Wu, and Ke Xu. Structural entropy guided unsupervised graph out-of-distribution detection. InProceedings of the AAAI Conference on Artificial Intelligence, AAAI-25, pages 17258–17266, 2025

  8. [16]

    Incremental measurement of structural entropy for dynamic graphs.Artificial Intelligence, 334:104175, 2024

    Runze Yang, Hao Peng, Chunyang Liu, and Angsheng Li. Incremental measurement of structural entropy for dynamic graphs.Artificial Intelligence, 334:104175, 2024

  9. [17]

    Yiwei Liu, Jiamou Liu, Zijian Zhang, Liehuang Zhu, and Angsheng Li. Rem: From structural entropy to community structure deception.Proceedings of the 33rd International Conference on Neural Information Processing Systems, NeurIPS-19, pages 12938–12948, 2019

  10. [18]

    Se-gsl: A general and effective graph structure learning framework through structural entropy optimization

    Dongcheng Zou, Hao Peng, Xiang Huang, Renyu Yang, Jianxin Li, Jia Wu, Chunyang Liu, and Philip S Yu. Se-gsl: A general and effective graph structure learning framework through structural entropy optimization. In Proceedings of the ACM Web Conference 2023, WWW-23, pages 499–510, 2023

  11. [19]

    Sega: Structural entropy guided anchor view for graph contrastive learning

    Junran Wu, Xueyuan Chen, Bowen Shi, Shangzhe Li, and Ke Xu. Sega: Structural entropy guided anchor view for graph contrastive learning. InProceedings of the 40th International Conference on Machine Learning, ICML-23, pages 37293–37312, 2023

  12. [20]

    A simple yet effective method for graph classification

    Junran Wu, Shangzhe Li, Jianhao Li, Yicheng Pan, and Ke Xu. A simple yet effective method for graph classification. InProceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI- 22, pages 3580–3586, 2022. 15 arXivTemplateA PREPRINT

  13. [21]

    Structural entropy guided graph hierarchical pooling

    Junran Wu, Xueyuan Chen, Ke Xu, and Shangzhe Li. Structural entropy guided graph hierarchical pooling. In Proceedings of the 39th International Conference on Machine Learning, ICML-22, pages 24017–24030, 2022

  14. [22]

    Guangjie Zeng, Hao Peng, Angsheng Li, Jia Wu, Chunyang Liu, and Philip S. Yu. Scalable semi-supervised clustering via structural entropy with different constraints.IEEE Transactions on Knowledge and Data Engineering, 37(1):478–492, 2025

  15. [23]

    Resistance maximization principle for defending networks against virus attack.Physica A: Statistical Mechanics and its Applications, 466:211–223, 2017

    Angsheng Li, Xiaohui Zhang, and Yicheng Pan. Resistance maximization principle for defending networks against virus attack.Physica A: Statistical Mechanics and its Applications, 466:211–223, 2017

  16. [24]

    Achieving graph clustering privacy preservation based on structure entropy in social iot.IEEE Internet of Things Journal, 9(4):2761–2777, 2022

    Youliang Tian, Zhiying Zhang, Jinbo Xiong, Lei Chen, et al. Achieving graph clustering privacy preservation based on structure entropy in social iot.IEEE Internet of Things Journal, 9(4):2761–2777, 2022

  17. [25]

    Hao Peng, Jingyun Zhang, Xiang Huang, Zhifeng Hao, Angsheng Li, Zhengtao Yu, and Philip S. Yu. Unsupervised social bot detection via structural information theory.ACM Transactions on Information Systems, 42(6):1–42, 2024

  18. [26]

    Yuwei Cao, Hao Peng, Zhengtao Yu, and Philip S. Yu. Hierarchical and incremental structural entropy minimization for unsupervised social event detection.Proceedings of the AAAI Conference on Artificial Intelligence, AAAI-24, pages 8255–8264, 2024

  19. [27]

    Sebot: Structural entropy guided multi-view contrastive learning for social bot detection

    Yingguang Yang, Qi Wu, Buyun He, Hao Peng, Renyu Yang, Zhifeng Hao, and Yong Liao. Sebot: Structural entropy guided multi-view contrastive learning for social bot detection. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD-24, page 38...

  20. [28]

    Valence-based homophily on twitter: Network analysis of emotions and political talk in the 2012 presidential election.New media & society, 18(7):1382–1400, 2016

    Itai Himelboim, Kaye D Sweetser, Spencer F Tinkham, Kristen Cameron, Matthew Danelo, and Kate West. Valence-based homophily on twitter: Network analysis of emotions and political talk in the 2012 presidential election.New media & society, 18(7):1382–1400, 2016

  21. [29]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, et al. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (lo...

  22. [30]

    Open graph benchmark: datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: datasets for machine learning on graphs. InProceedings of the 34th International Conference on Neural Information Processing Systems, NeurIPS-...

  23. [31]

    Harnessing explanations: LLM-to-LM interpreter for enhanced text-attributed graph representation learning

    Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, and Bryan Hooi. Harnessing explanations: LLM-to-LM interpreter for enhanced text-attributed graph representation learning. InThe Twelfth International Conference on Learning Representations, ICLR-24, 2024

  24. [32]

    Can gnn be good adapter for llms? InProceedings of the ACM Web Conference 2024, WWW-24, page 893–904, 2024

    Xuanwen Huang, Kaiqiao Han, Yang Yang, Dezheng Bao, Quanjin Tao, Ziwei Chai, and Qi Zhu. Can gnn be good adapter for llms? InProceedings of the ACM Web Conference 2024, WWW-24, page 893–904, 2024

  25. [33]

    Cohen, and Ruslan Salakhutdinov

    Zhilin Yang, William W. Cohen, and Ruslan Salakhutdinov. Revisiting semi-supervised learning with graph embeddings. InProceedings of the 33rd International Conference on International Conference on Machine Learning, ICML-16, page 40–48, 2016

  26. [34]

    Taglas: An atlas of text-attributed graph datasets in the era of large graph and language models, 2024

    Jiarui Feng, Hao Liu, Lecheng Kong, Yixin Chen, and Muhan Zhang. Taglas: An atlas of text-attributed graph datasets in the era of large graph and language models, 2024

  27. [35]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, ICLR-17, 2017

  28. [36]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. InInternational Conference on Learning Representations, ICLR-18, 2018

  29. [37]

    How powerful are graph neural networks? In International Conference on Learning Representations, ICLR-19, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations, ICLR-19, 2019

  30. [38]

    Inductive representation learning on large graphs.Proceedings of the 31st International Conference on Neural Information Processing Systems, NeurIPS-17, 30, 2017

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs.Proceedings of the 31st International Conference on Neural Information Processing Systems, NeurIPS-17, 30, 2017

  31. [39]

    Degree-based stratification of nodes in graph neural networks

    Ameen Ali, Lior Wolf, and Hakan Cevikalp. Degree-based stratification of nodes in graph neural networks. In Proceedings of the 15th Asian Conference on Machine Learning, volume 222, pages 15–27. PMLR, 2024

  32. [40]

    Graph-skeleton:˜ 1% nodes are sufficient to represent billion-scale graph

    Linfeng Cao, Haoran Deng, Yang Yang, Chunping Wang, and Lei Chen. Graph-skeleton:˜ 1% nodes are sufficient to represent billion-scale graph. InProceedings of the ACM Web Conference 2024, WWW-24, pages 570–581, 2024

  33. [41]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024. 16 arXivTemplateA PREPRINT

  34. [42]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. InProceedings of the 37th International Conference on Neural Information Processing Systems, NeurIPS-23, 2023

  35. [43]

    Multi-scale heterogeneous text-attributed graph datasets from diverse domains

    Yunhui Liu, Qizhuo Xie, Jinwei Shi, et al. Multi-scale heterogeneous text-attributed graph datasets from diverse domains. InCompanion Proceedings of the ACM on Web Conference 2025, WWW-25, pages 757–760, 2025

  36. [44]

    A survey of structural entropy: Theory, methods, and applications

    Dingli Su, Hao Peng, Yicheng Pan, and Angsheng Li. A survey of structural entropy: Theory, methods, and applications. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25, pages 10660–10668, 2025. Survey Track

  37. [45]

    Item recommendation on monotonic behavior chains

    Mengting Wan and Julian McAuley. Item recommendation on monotonic behavior chains. InProceedings of the 12th ACM Conference on Recommender Systems, RecSys-18, page 86–94, 2018

  38. [46]

    Oag-bench: a human-curated benchmark for academic graph mining

    Fanjin Zhang, Shijie Shi, et al. Oag-bench: a human-curated benchmark for academic graph mining. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD-24, pages 6214–6225, 2024

  39. [47]

    Explainability in graph neural networks: A taxonomic survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5782–5799, 2023

    Hao Yuan, Haiyang Yu, Shurui Gui, and Shuiwang Ji. Explainability in graph neural networks: A taxonomic survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5782–5799, 2023

  40. [48]

    Moleculenet: a benchmark for molecular machine learning.Chemical science, 9(2):513–530, 2018

    Zhenqin Wu, Bharath Ramsundar, Evan N Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S Pappu, Karl Leswing, and Vijay Pande. Moleculenet: a benchmark for molecular machine learning.Chemical science, 9(2):513–530, 2018

  41. [49]

    Graph coarsening via convolution matching for scalable graph neural network training

    Charles Dickens, Edward Huang, Aishwarya Reganti, Jiong Zhu, Karthik Subbian, and Danai Koutra. Graph coarsening via convolution matching for scalable graph neural network training. InCompanion Proceedings of the ACM Web Conference 2024, WWW-24, pages 1502–1510, 2024

  42. [50]

    Featured graph coarsening with similarity guarantees

    Manoj Kumar, Anurag Sharma, Shashwat Saxena, and Sandeep Kumar. Featured graph coarsening with similarity guarantees. InProceedings of the 40th International Conference on Machine Learning, ICML-23, pages 17953– 17975, 2023

  43. [51]

    Ugc: universal graph coarsening

    Mohit Kataria, Sandeep Kumar, and Jayadeva. Ugc: universal graph coarsening. InProceedings of the 38th International Conference on Neural Information Processing Systems, NeurIPS-24, pages 63057–63081, 2024

  44. [52]

    Quantifying node importance over network structural stability

    Fan Zhang, Qingyuan Linghu, Jiadong Xie, Kai Wang, Xuemin Lin, and Wenjie Zhang. Quantifying node importance over network structural stability. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD-23, page 3217–3228, 2023. 17 arXivTemplat...

  45. [53]

    The collective semantic meaning

  46. [54]

    Answer: Box 5The Aggregation prompt template for Molecule Property Prediction Datasets

    Overall sentiment tendency (Positive/Negative). Answer: Box 5The Aggregation prompt template for Molecule Property Prediction Datasets. The Prompt Template for Molecule Dataset. Molecule Template: Analyze the following atoms in a molecular substructure:{RAW_TEXTS} Task: Descri...

  47. [55]

    The functional group or structural motif

  48. [56]

    is similar to

    Any notable chemical properties. Answer: 8.5.2 Downstream Task Prompt Template Box 6The Downstream Node-level Task prompt template. The Prompt Template for Node-level classification. Classification Template:Given a citation graph, the 0th node (target paper) has the following ...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.