Pith. sign in

REVIEW 4 major objections 5 minor 45 references

UTCS: Effective Unsupervised Temporal Community Search with Pre-training of Temporal Dynamics and Subgraph Knowledge

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Temporal community search works without labels: UTCS beats baselines by 60% F1 on average.

desk verdict Workmanlike new combination for unsupervised temporal community search; credible results, but missing definitions and an overclaimed headline metric need fixing. read the letter →

arxiv 2506.02784 v1 pith:4VSVHFGT submitted 2025-06-03 cs.IR

Yue Zhang , Yankai Chen , Yingli Zhou , Yucan Guo , Xiaolin Han , Chenhao Ma This is my paper
classification cs.IR
keywords communitysearchtemporalgraphunsupervisedlearningHawkesprocesspre-trainingLeidenalgorithmnodeembeddingcontrastive
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

The paper tries to establish that temporal community search—finding the cohesive subgraph around a query node when edges carry timestamps—can be solved by a purely unsupervised learning method. It proposes UTCS, which pre-trains node embeddings on the temporal graph without labels, then searches locally for the community around the query. The reported experiments claim the method beats all compared baselines on five real-world datasets, with an average F1-score improvement of 60.44% over the second-best model. If true, this would remove the need for costly labeled data and for hand-defined subgraph patterns like k-core in temporally evolving networks.

What carries the argument

The machinery is a two-phase unsupervised pipeline: an offline pre-training objective $L = L_{\text{tmp}} + L_{\text{node}} + L_{\text{batch}}$, and an online local search. The temporal dynamics learner uses a conditional intensity function $\lambda(u,v,t)$ from the Hawkes process to model the exponentially decaying influence of past interactions; the node-Leiden alignment uses a Student's t-distribution assignment $q$ and a squared normalized target $p$, minimized by KL divergence; the batch refinement uses a contrastive cosine-similarity loss over historical neighbors. The online search builds a candidate set $D$ from Leiden subgraphs on the de-temporal graph and their top-$k$ similar subgraphs, then greedily adds the highest-scoring node while ECSG increases. This combination lets the model use temporal interaction patterns and graph structure jointly without any supervision.

What would settle it

Synthesize a temporal graph where the ground-truth community around the query nodes is connected only through time-respecting paths that pass through outside nodes, so the community has no static edges in the de-temporal graph. If UTCS still returns the ground-truth community, the paper's candidate-set assumption is violated; if it fails, the assumption is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that UTCS is the first learning-based temporal community search method that needs no labeled data and no predefined subgraph structure. UTCS captures temporal dynamics through a Hawkes-process likelihood term that treats past interactions as boosting future ones, aligns node embeddings with Leiden-community assignments using Student's t-distribution and KL divergence, and refines embeddings by reconstructing adjacency in batches. At query time, it restricts attention to the Leiden subgraphs containing the query nodes plus their k nearest neighbors, scores nodes by cosine similarity to the average query embedding, and grows the community greedily while the Expected Community Score Gain increases. On five datasets the paper reports UTCS as consistently best, with 60.44% average F1 improvement over the second-best baseline.

Load-bearing premise

The load-bearing premise is that the true temporal community is visible as a static subgraph in the de-temporal graph, since the candidate set is built only from Leiden subgraphs of the graph with timestamps removed; a community that is temporally coherent but structurally absent there can never be retrieved.

Editorial extensions

If this is right

  • Temporal community search can be performed on graphs without any labeled ground truth, removing the annotation bottleneck for new domains.
  • The offline pre-training is reusable: once trained, the same embedding network can answer online queries on the same graph, and the local search design keeps per-query cost low.
  • The model avoids predefined structures like k-core or k-truss, so it can fit communities whose topology does not satisfy those strict constraints.
  • Compared to baselines, UTCS is reported to scale to large graphs where two static learning baselines run out of memory, and to run over 100× faster than the query-centered baseline on the School dataset.

Reading between the lines

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

  • The candidate-set construction could be the main recall bottleneck; a temporal-aware partitioning that respects timestamps might recover communities that Leiden on the de-temporal graph misses.
  • The 60.44% average improvement is driven partly by datasets where several baselines fail with out-of-memory or return empty results; on datasets where all baselines run, the margin is smaller.
  • The same pre-training objectives could be adapted to supervised or semi-supervised temporal community search, or to node classification on temporal graphs, since the Hawkes-process term is task-agnostic.
  • One could test the sensitivity of the method to the hyperparameter $k$ (number of similar subgraphs) and to the Leiden resolution; the current paper fixes $k=2$ and does not report that sensitivity.
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 / 5 minor

Summary. The paper proposes UTCS, an unsupervised learning-based method for temporal community search. It pre-trains node embeddings with three objectives: a temporal Hawkes-process loss (Eq. 1), a Node-Leiden subgraph alignment loss based on Student's t-distribution and KL divergence (Eq. 2), and a batch-level contrastive refinement loss (Eq. 3). In the online search phase, Algorithm 1 builds a candidate set from Leiden communities of the de-temporal graph that contain the query nodes plus their top-k similar subgraphs, scores candidate nodes by cosine similarity to the query embedding, and greedily adds the highest-scoring node while the Expected Community Score Gain (ECSG) increases. Experiments on five real-world datasets report F1, Jaccard, and NMI against several baselines, with an ablation study and efficiency comparisons. The paper claims a 60.44% average F1 improvement over COCLEP and positions UTCS as the first unsupervised learning-based temporal community search method.

Significance. If the claimed results hold, UTCS would be a valuable contribution: it is a plausible first step toward label-free temporal community search, and the design of combining temporal dynamics with structural subgraph knowledge is reasonable. The paper ships code and includes ablations that attribute much of the gain to temporal modeling, which is a strength. However, the significance is currently undercut by three load-bearing problems: (1) the candidate generation step in Algorithm 1 can exclude the true community before any scoring occurs, and this bottleneck is not analyzed; (2) two central algorithmic components (ECSG and TopKSimC) and the Hawkes conditional intensity are never defined, making the method irreproducible from the paper; and (3) the headline 60.44% improvement is computed against COCLEP, which is not the second-best method on several datasets, and the claim that UTCS "consistently achieves the best results" is contradicted by Table 2 on the Patent dataset. These issues are fixable but currently prevent an objective assessment of the method's true standing.

major comments (4)
  1. [3.3, Algorithm 1] The candidate set D is constructed in lines 1-4 as the union of Leiden communities of the de-temporal graph G' that contain the query nodes plus their k most similar subgraphs. Since all scoring and the final greedy selection are restricted to D (lines 5-14), the recall of D is an upper bound on the final community's recall. The paper provides no analysis of recall@D, no sensitivity study with respect to Leiden resolution, and no comparison against global scoring on small datasets. A temporal community whose aggregated static edges are not structurally dense may be split across or absent from these Leiden clusters, and then no pre-trained embedding or scoring mechanism can recover it. This matters because the central claim that temporal modeling drives the gains is not substantiated: the ablation study (Table 3) removes the Node-Leiden alignment loss but still uses the same Leiden partition for candidate generation, so it does not isolate the effect of the candidate bottleneck. Please report recall@D for each dataset, run a variant that scores all nodes globally, and vary the Leiden resolution or community-detection algorithm to show that the method is robust to the candidate construction.
  2. [3.2.1, Eq. (1)] The temporal loss in Eq. (1) depends on the conditional intensity function λ(u,v,t) of the Hawkes process, but this function is never defined. Without a precise specification (e.g., an exponential kernel over past events with learnable parameters, as in the cited works), the loss L_tmp is not reproducible and it is impossible to verify that the model actually captures temporal dynamics. Similarly, Eq. (2) introduces the assignment probability q'(v,i,t) without specifying how it is computed from the embeddings beyond referencing Student's t-distribution; the exact form and the role of the updated embeddings need to be stated. Please provide complete definitions of all quantities in Eqs. (1)-(3), including the intensity function and the normalization constant.
  3. [3.3, Algorithm 1] The pseudo-code uses two components that are never defined in the text: ECSG (Expected Community Score Gain) and TopKSimC. ECSG is only described verbally as 'the gap between the scores of community nodes and the expected scores of random nodes' with a citation to TransZero, and TopKSimC is not described at all. As a result, the online search phase cannot be implemented from the paper alone, and the reported efficiency and effectiveness results cannot be independently verified. Please give explicit mathematical definitions of ECSG(S, T, G') and of the similarity measure and retrieval procedure used in TopKSimC.
  4. [4.2, Table 2] The claim that UTCS 'consistently achieves the best results' and improves F1-score by 60.44% over 'the second-best model, COCLEP' is not supported by the table. On the Patent dataset, UTCS obtains F1=0.31, while QD-GNN obtains 0.38, making UTCS third-best, not first. Furthermore, COCLEP is not the second-best method on Brain (TransZero is) or arXivAI (TransZero is) or arXivCS (TransZero is, with COCLEP second). The 60.44% figure appears to be the arithmetic mean of per-dataset relative improvements of UTCS over COCLEP only, not over the actual second-best method. This averaging is misleading because it ignores stronger baselines on several datasets. Please report the per-dataset improvements over the per-dataset best and second-best methods, state the exact formula for the average, and correct the claim of consistent superiority.
minor comments (5)
  1. [1, Contribution bullet 3] The phrase 'an average improvement of 60.44% in F1-score compared to the latest competitors' is vague; please specify which competitors and which averaging scheme are used.
  2. [2, Problem Formulation] Problem 1 defines the TCS problem only as 'return a subgraph containing Q'. Since there is no notion of ground truth or desired objective, it would be helpful to state what constitutes a correct output or how evaluation is performed with respect to ground-truth communities.
  3. [4.1, Table 2] The table formatting is inconsistent: 'QD-GNN0.75' lacks a space, the 'Rank' column (e.g., '7/7/7') is not explained, and 'OOM' and '-' are used without a clear legend. Please clean up the presentation and define all symbols.
  4. [4.1, Implementation Details] Table 3 reports F1 values without standard deviations, although the text states results are averaged over five runs. Please include variance or state why it is omitted.
  5. [4.2, Ablation Study] The ablation variants are not described precisely: it is unclear what 'UTCS-BR-TM' removes (apparently both Batch Refinement and Temporal Modeling) and why the naming is inconsistent (e.g., 'UTCS-TM' vs 'UTCS-NA'). Please clarify the naming and the exact modules removed in each variant.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation found; minor self-citations are not load-bearing.

full rationale

The paper is an empirical method paper; we find no step in which a claimed prediction is equivalent, by construction, to a fitted input or to a self-citation. The only self-referential mechanism is Node-Leiden Subgraph Alignment: Leiden partitions of the de-temporal graph G' are used as pseudo-labels in the pretraining loss (Eq. 2) and again to build the candidate set D in Algorithm 1 (lines 1-4). This is a standard self-training loop, not a parameter fitted to the target metric and then renamed a prediction; the final evaluation is against external ground-truth communities. ECSG and TopKSimC are invoked from [31] and left undefined in the paper, which is a reproducibility defect rather than circularity. The paper's self-citations (e.g., refs. [3], [4], [9], [38]) are background references to prior community-search work by co-authors and are not load-bearing for the UTCS construction or its empirical claims. Score 2 reflects the presence of minor non-load-bearing self-citations; no circular derivation was found.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The ledger is light on free parameters for a neural model because the paper makes no theoretical derivation; all listed parameters are training hyperparameters. The main burden lies in domain assumptions: the data are assumed to follow a Hawkes process, static Leiden structure is assumed to reveal temporal communities, and the ECSG criterion from [31] is imported without justification. No new entities are invented.

free parameters (8)
  • Negative sample count l = 3
    Used in the temporal loss Eq. (1) to sample negative nodes; chosen by hand in Section 4.1.
  • Historical neighbor count h = 3
    Used in batch-level refinement Eq. (3) to select positive samples; set in Section 4.1.
  • Temperature T = 0.5
    Controls sharpness of cosine similarity in Eq. (3); chosen in Section 4.1.
  • Top-k similar subgraphs k = 2
    Controls the number of similar Leiden subgraphs added to the candidate set in Algorithm 1 line 3.
  • Embedding dimension = 128
    Node embedding size; chosen in Section 4.1.
  • Learning rate = 0.01
    Training hyperparameter listed in Section 4.1.
  • Batch size = 1024
    Training hyperparameter listed in Section 4.1.
  • Number of epochs = 200
    Training hyperparameter listed in Section 4.1.
assumptions (5)
  • domain assumption Temporal interactions follow a Hawkes process with exponentially decaying influence.
    The temporal loss Eq. (1) assumes past edges boost the conditional intensity of future events; if the data do not follow this process, the learned representations may not capture temporal structure.
  • domain assumption Leiden algorithm on the de-temporal graph produces subgraphs that correspond to meaningful communities.
    The pseudo-labels in Section 3.2.2 and the online candidate set in Algorithm 1 both depend on Leiden partitioning of G'; communities that are not structurally dense in the static projection may be missed.
  • domain assumption The ECSG criterion from TransZero [31] is a valid measure of community quality.
    The greedy expansion in Algorithm 1 uses ECSG to decide when to stop; the paper does not re-derive or justify it for temporal graphs.
  • domain assumption The de-temporal graph G' preserves enough information for candidate generation despite removing timestamps.
    If temporal communities are not reflected in the static merged graph, the candidate search space may exclude the true community.
  • ad hoc to paper Student's t-distribution and squared-normalized targets improve cluster alignment.
    This DEC-style self-training heuristic in Section 3.2.2 is introduced without theoretical justification; it is a common but ad hoc technique.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UTCS: Effective Unsupervised Temporal Community Search with Pre-training of Temporal Dynamics and Subgraph Knowledge." pith.science (2026). https://pith.science/paper/4VSVHFGT

@misc{pith2026250602784,
  author       = {Pith},
  title        = {Pith review of: UTCS: Effective Unsupervised Temporal Community Search with Pre-training of Temporal Dynamics and Subgraph Knowledge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VSVHFGT}},
  note         = {Machine review of arXiv:2506.02784}
}
read the original abstract

In many real-world applications, the evolving relationships between entities can be modeled as temporal graphs, where each edge has a timestamp representing the interaction time. As a fundamental problem in graph analysis, {\it community search (CS)} in temporal graphs has received growing attention but exhibits two major limitations: (1) Traditional methods typically require predefined subgraph structures, which are not always known in advance. (2) Learning-based methods struggle to capture temporal interaction information. To fill this research gap, in this paper, we propose an effective \textbf{U}nsupervised \textbf{T}emporal \textbf{C}ommunity \textbf{S}earch with pre-training of temporal dynamics and subgraph knowledge model (\textbf{\model}). \model~contains two key stages: offline pre-training and online search. In the first stage, we introduce multiple learning objectives to facilitate the pre-training process in the unsupervised learning setting. In the second stage, we identify a candidate subgraph and compute community scores using the pre-trained node representations and a novel scoring mechanism to determine the final community members. Experiments on five real-world datasets demonstrate the effectiveness.

Figures

Figures reproduced from arXiv: 2506.02784 by the authors.

Figure 1
Figure 1. The architecture of our model. existing learning-based approach for temporal community search in the unsupervised setting. Our solution. To address the aforementioned limitations, we have developed an effective model, namely Unsupervised Temporal Community Search with pre-training of temporal dynamics and subgraph knowledge (UTCS). UTCS incorporates temporal and structural information at both local and global scales… view at source ↗
Figure 3
Figure 3. A case study on arXivAI dataset. on the arXivAI dataset, where QD-GNN and CS-TGN encountered OOM issues. In [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Efficiency results of CS methods. methods in the search phase, with OOM denoting out-of-memory. Our model outperforms the baselines in terms of efficiency on most datasets, achieving over 100× speedup on the School dataset compared to QTCS due to our local search design. While CS-TGN is slightly more efficient than ours on the Brain and Patent datasets, it faces effectiveness issues and struggles to scale on large g… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 33 canonical work pages

  1. [31]

    Jianwei Wang, Kai Wang, Xuemin Lin, Wenjie Zhang, and Ying Zhang. 2024. Efficient Unsupervised Community Search with Pre-trained Graph Transformer. arXiv:2403.18869 [cs]

    arXiv 2024

  2. [1]

    Esra Akbas and Peixiang Zhao. 2017. Truss-based community search: a truss- equivalence based indexing approach. VLDB 10, 11 (2017), 1298–1309

    work page 2017

  3. [2]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations (ICML’20). JMLR.org, Article 149, 11 pages

    work page 2020

  4. [3]

    Yankai Chen, Yixiang Fang, Reynold Cheng, Yun Li, Xiaojun Chen, and Jie Zhang

    work page

  5. [4]

    Yankai Chen, Jie Zhang, Yixiang Fang, Xin Cao, and Irwin King. 2021. Effi- cient community search over large directed graphs: An augmented index-based approach. In Proceedings of the Twenty-Ninth International Conference on Interna- tional Joint Conferences on Artificial Intelligence . 3544–3550

    work page 2021

  6. [5]

    Lingyang Chu, Yanyan Zhang, Yu Yang, Lanjun Wang, and Jian Pei. 2019. Online density bursting subgraph detection from temporal graphs. Proceedings of the VLDB Endowment 12, 13 (2019), 2353–2365

    work page 2019

  7. [6]

    Wanyun Cui, Yanghua Xiao, Haixun Wang, Yiqi Lu, and Wei Wang. 2013. Online search of overlapping communities. In Proceedings of the 2013 ACM SIGMOD international conference on Management of data . 277–288

    2013

  8. [7]

    Wanyun Cui, Yanghua Xiao, Haixun Wang, and Wei Wang. 2014. Local search of communities in large graphs. In SIGMOD. 991–1002

    work page 2014

Show all 45 references
  1. [8]

    Leon Danon, Albert Diaz-Guilera, Jordi Duch, and Alex Arenas. 2005. Comparing community structure identification. Journal of statistical mechanics: Theory and experiment 2005, 09 (2005), P09008

    2005

  2. [9]

    Yixiang Fang, Reynold Cheng, Yankai Chen, Siqiang Luo, and Jiafeng Hu. 2017. Effective and efficient attributed community search. The VLDB Journal 26 (2017), 803–828

    2017

  3. [10]

    Yixiang Fang, Xin Huang, Lu Qin, Ying Zhang, Wenjie Zhang, Reynold Cheng, and Xuemin Lin. 2019. A survey of community search over big graphs.The VLDB Journal 29, 1 (July 2019), 353–392. https://doi.org/10.1007/s00778-019-00556-x

    2019 doi

  4. [11]

    Jun Gao, Jiazun Chen, Zhao Li, and Ji Zhang. 2021. ICS-GNN: Lightweight Interactive Community Search via Graph Neural Network. Proceedings of the VLDB Endowment 14, 6 (Feb. 2021), 1006–1018. https://doi.org/10.14778/3447689. 3447704

    2021 doi

  5. [12]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. In SIGKDD. 855–864

    2016

  6. [13]

    Bronwyn H Hall, Adam B Jaffe, and Manuel Trajtenberg. 2001. The NBER patent citation data file: Lessons, insights and methodological tools

    2001

  7. [14]

    Farnoosh Hashemi, Ali Behrouz, and Milad Rezaei Hajidehi. 2023. Cs-tgn: Com- munity search via temporal graph neural networks. In WebConf. 1196–1203

    2023

  8. [15]

    Alan G Hawkes. 1971. Spectra of some self-exciting and mutually exciting point processes. Biometrika 58, 1 (1971), 83–90

    1971

  9. [16]

    Xin Huang, Laks VS Lakshmanan, Jeffrey Xu Yu, and Hong Cheng. 2015. Approx- imate closest community search in networks. arXiv preprint arXiv:1505.05956 (2015)

    2015 arXiv

  10. [17]

    Yuli Jiang, Yu Rong, Hong Cheng, Xin Huang, Kangfei Zhao, and Junzhou Huang

  11. [18]

    Solomon Kullback and Richard A Leibler. 1951. On information and sufficiency. The annals of mathematical statistics 22, 1 (1951), 79–86

    1951

  12. [19]

    Ling Li, Siqiang Luo, Yuhai Zhao, Caihua Shan, Zhengkui Wang, and Lu Qin

  13. [20]

    Longlong Lin, Pingpeng Yuan, Rong-Hua Li, Chunxue Zhu, Hongchao Qin, Hai Jin, and Tao Jia. 2024. QTCS: Efficient Query-Centered Temporal Community Search. Proceedings of the VLDB Endowment 17, 6 (2024), 1187–1199

    2024

  14. [21]

    Meng Liu, Yue Liu, Ke Liang, Wenxuan Tu, Siwei Wang, Sihang Zhou, and Xinwang Liu. 2024. DEEP TEMPORAL GRAPH CLUSTERING. (2024)

    2024

  15. [22]

    Wensheng Luo, Yixiang Fang, Chunxu Lin, and Yingli Zhou. 2024. Efficient Parallel D-Core Decomposition at Scale. Proceedings of the VLDB Endowment 17, 10 (2024), 2654–2667

    2024

  16. [23]

    Wensheng Luo, Qiaoyuan Yang, Yixiang Fang, and Xu Zhou. 2023. Efficient core maintenance in large bipartite graphs. Proceedings of the ACM on Management of Data 1, 3 (2023), 1–26

    2023

  17. [24]

    Rossana Mastrandrea, Julie Fournet, and Alain Barrat. 2015. Contact patterns in a high school: a comparison between data collected using wearable sensors, contact diaries and friendship surveys. PloS one 10, 9 (2015), e0136497

    2015

  18. [25]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. NeurIPS 26 (2013)

    2013

  19. [26]

    Maria Giulia Preti, Thomas AW Bolton, and Dimitri Van De Ville. 2017. The dynamic functional connectome: State-of-the-art and perspectives. Neuroimage 160 (2017), 41–54

    2017

  20. [27]

    Hongchao Qin, Rong-Hua Li, Guoren Wang, Lu Qin, Yurong Cheng, and Ye Yuan. 2019. Mining periodic cliques in temporal networks. In 2019 IEEE 35th International Conference on Data Engineering (ICDE) . IEEE, 1130–1141

    2019

  21. [28]

    Mauro Sozio and Aristides Gionis. 2010. The community-search problem and how to plan a successful cocktail party. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining . 939–948

    2010

  22. [29]

    Vincent A Traag, Ludo Waltman, and Nees Jan Van Eck. 2019. From Louvain to Leiden: guaranteeing well-connected communities. Scientific reports (2019)

    2019

  23. [30]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  24. [32]

    Jianwei Wang, Kai Wang, Xuemin Lin, Wenjie Zhang, and Ying Zhang. 2024. Neural Attributed Community Search at Billion Scale. arXiv:2403.18874 [cs]

    2024 arXiv

  25. [33]

    Kuansan Wang, Zhihong Shen, Chiyuan Huang, Chieh-Han Wu, Yuxiao Dong, and Anshul Kanakia. 2020. Microsoft academic graph: When experts are not enough. Quantitative Science Studies 1, 1 (2020), 396–413

    2020

  26. [34]

    Long Yuan, Lu Qin, Wenjie Zhang, Lijun Chang, and Jianye Yang. 2017. Index- based densest clique percolation community search in networks. TKDE 30, 5 (2017), 922–935

    2017

  27. [35]

    Yuanyuan Zeng, Yixiang Fang, Chenhao Ma, Xu Zhou, and Kenli Li. 2024. Efficient distributed hop-constrained path enumeration on large-scale graphs. Proceedings of the ACM on Management of Data 2, 1 (2024), 1–25

    2024

  28. [36]

    Yuanyuan Zeng, Chenhao Ma, and Yixiang Fang. 2024. Distributed Shortest Distance Labeling on Large-Scale Graphs. Proceedings of the VLDB Endowment 17, 10 (2024), 2641–2653

    2024

  29. [37]

    Yao Zhang, Yun Xiong, Yun Ye, Tengfei Liu, Weiqiang Wang, Yangyong Zhu, and Philip S Yu. 2020. SEAL: Learning heuristics for community detection with generative adversarial networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data M...

    2020

  30. [38]

    Yifei Zhang, Hao Zhu, Zixing Song, Yankai Chen, Xinyu Fu, Ziqiao Meng, Pi- otr Koniusz, and Irwin King. 2024. Geometric view of soft decorrelation in self-supervised learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4338–4349

    2024

  31. [39]

    Yingli Zhou, Yixiang Fang, Wensheng Luo, and Yunming Ye. 2023. Influential community search over large heterogeneous information networks. VLDB 16, 8 (2023), 2047–2060

    2023

  32. [40]

    Yingli Zhou, Yixiang Fang, Chenhao Ma, Tianci Hou, and Xin Huang. 2024. Effi- cient Maximal Motif-Clique Enumeration over Large Heterogeneous Information Networks. Proceedings of the VLDB Endowment 17, 11 (2024), 2946–2959

    2024

  33. [41]

    Yingli Zhou, Qingshuo Guo, Yixiang Fang, and Chenhao Ma. 2024. A Counting- based Approach for Efficient k-Clique Densest Subgraph Discovery. Proceedings of the ACM on Management of Data 2, 3 (2024), 1–27

    2024

  34. [42]

    Yuan Zuo, Guannan Liu, Hao Lin, Jia Guo, Xiaoqian Hu, and Junjie Wu. 2018. Embedding Temporal Network via Neighborhood Formation. In SIGKDD. 2857– 2866

    2018

  35. [2018]

    IEEE Transactions on Knowledge and Data Engineering 31, 8 (2018), 1624–1629

    Exploring communities in large profiled graphs. IEEE Transactions on Knowledge and Data Engineering 31, 8 (2018), 1624–1629

    2018

  36. [2022]

    arXiv:2104.03583 [cs]

    Query Driven-Graph Neural Networks for Community Search: From Non- Attributed, Attributed, to Interactive Attributed. arXiv:2104.03583 [cs]

    arXiv

  37. [2023]

    Contrastive Learning-based Semi-Supervised Community Search. In ICDE

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.