Pith. sign in

REVIEW 2 major objections 4 minor 21 references

The paper argues that directed dynamic graph edges are classified more accurately when source and destination are modeled as separate roles and aligned by a self-supervised contrastive pretraining step that uses the full unlabeled interacti

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 20:26 UTC pith:2FVIHD7R

load-bearing objection DyGnROLE is a genuine architectural contribution, but the headline SOTA claim rests on an unfair comparison: pretraining access is confounded with architecture, so the reported margins are not attributable to role-aware modeling alone. the 2 major comments →

arxiv 2602.23135 v2 pith:2FVIHD7R submitted 2026-02-26 cs.LG cs.AIcs.SI

DyGnROLE: Asymmetric Pretraining for Edge Classification on Dynamic Graphs

Tyler Bonnet , Marek Rei This is my paper
classification cs.LG cs.AIcs.SI
keywords dynamic graphsedge classificationsource-destination asymmetryrole-aware representation learningself-supervised pretrainingtemporal contrastive learninglow-label learningcontinuous-time dynamic graphs
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most dynamic-graph models treat every node the same way whether it is initiating or receiving an interaction. The paper tries to establish that this symmetry is a real cost: when the same node acts as source versus destination it occupies a different behavioral role, and edge classification improves when the architecture encodes that distinction from the start. The proposed model gives sources and destinations separate embedding streams and role-specific positional signals, then aligns those separated spaces with a self-supervised pretraining objective over the unlabeled interaction history. Evaluated with only 10,000 labeled edges, it reaches the best Macro F1 on seven of eight datasets and ties on the eighth, while several baselines collapse to predicting the majority class. The sympathetic reading is that the paper makes role-awareness and asymmetric pretraining load-bearing for low-label dynamic-graph learning.

Core claim

On its own terms, DyGnROLE's central claim is that role-aware modeling is not a small auxiliary feature but the main mechanism that lets a dynamic-graph Transformer generalize under label scarcity. It separates source and destination identities into distinct learnable vocabularies, injects four role-semantic positional identifiers before self-attention, and pools each side with its own CLS token. The new pretraining task, Temporal Contrastive Link Prediction (TCLP), treats the source-destination alignment as a retrieval problem and, crucially, masks all historically observed pairs from negatives so the training signal follows the temporal direction in which pairs become known. Across eight d

What carries the argument

The load-bearing mechanism is the combination of (1) separate source/destination embedding vocabularies plus four learnable role embeddings added to query and neighbor positions, so the Transformer cannot collapse the two roles; (2) within- and cross-sequence frequency counts that encode directed structural motifs; (3) dual-CLS pooling that yields distinct z_u and z_v representations; and (4) TCLP, a contrastive objective that pulls the normalized CLS vectors of matched source-destination pairs together while masking historical neighbors as false negatives. The masking is what makes the pretraining temporally directional: a previously observed pair is never penalized as a negative, only alig

Load-bearing premise

The headline numerical comparison assumes each baseline is trained only on the 10,000 labeled edges while the proposed model also sees the full unlabeled interaction history, so the gap could reflect data access rather than role-awareness.

What would settle it

Retrain the strongest baseline architecture under the identical protocol but with TCLP-style pretraining on the same unlabeled history; if it then matches or exceeds DyGnROLE's Macro F1 across the eight datasets, the claim that role-disentangled encoding is responsible is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • End-to-end future edge classification in low-label regimes becomes feasible: a model can learn from the abundant unlabeled interaction stream and be fine-tuned with as few as 10,000 annotated edges.
  • Architectural role separation should be considered a default design option for directed dynamic graphs, since shared-parameter baselines in the paper produce exactly symmetric representations for source and destination.
  • The TCLP masking rule can be reused as a temporal constraint in other settings where observed interactions should not be treated as negatives.
  • Reported gains are largest on datasets where baselines collapse to majority-class prediction, suggesting a clear diagnostic: if a baseline equals majority-class F1, it has not learned directional structure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next experiment is to give a strong symmetric baseline the same TCLP-style pretraining; if the gap shrinks to noise, the measured benefit comes from unlabeled-history access rather than role separation—the paper does not include that control.
  • The same role-disentanglement recipe should transfer to recommender systems and knowledge-graph completion, where user/item and subject/object are structurally distinct; testing there would show whether the result generalizes beyond the paper's eight datasets.
  • Because the ablation shows pretraining is the dominant driver, a simpler encoder paired with TCLP may capture much of the gain; the paper does not test how much architecture matters once the objective is fixed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes DyGnROLE, a Transformer-based continuous-time dynamic graph model that explicitly separates source and destination roles through distinct embedding tables, role-semantic positional encodings, within/cross-sequence neighbor frequency embeddings, and dual-CLS pooling. It also introduces TCLP, an InfoNCE-style pretraining objective with historical false-negative masking, and evaluates future edge classification on eight DTGB datasets in a 10,000-labeled-edge regime. The authors report the highest Macro F1 on 7 of 8 datasets and a tie on the 8th, with an ablation showing that all components, especially pretraining, contribute. An asymmetry-score analysis is used to claim that existing baselines are role-blind.

Significance. The architecture is clearly specified and the pretraining idea is sensible; if the comparisons were fair, the paper would demonstrate a useful recipe for label-efficient edge classification. Strengths include a precise model definition, a detailed hyperparameter appendix, use of public benchmark datasets, and an ablation with standard deviations. The main empirical claim, however, is currently not attributable to the role-aware components because the protocol gives DyGnROLE unlabeled pretraining that baselines do not receive. The asymmetry analysis in §5.3 is largely by construction. With additional controlled comparisons, the paper could make a solid contribution.

major comments (2)
  1. [§4.2 / Table 1 / Figure 2] The headline SOTA claim rests on an unequal comparison. DyGnROLE is pretrained with TCLP on the full training+validation interaction history, while every baseline is trained from scratch on the 10k labeled edges. The ablation in Figure 2 shows that removing pretraining collapses DyGnROLE to majority-class level (Stack U 53.9→44.6, Yelp 33.0→14.0). Thus the margins in Table 1 may reflect access to unlabeled data rather than role-disentangled encoding. Please add self-supervised dynamic-graph baselines (e.g., CLDG, IDOL, DyGMAE, DVGMAE) and/or pretrain a standard CTDGN with an equivalent generic contrastive or link-prediction objective on the same unlabeled data. Also report standard deviations or confidence intervals for Table 1, since several differences are small (GDELT 0.6 vs 0.4; ICEWS tie at 2.1).
  2. [§5.3 / Table 2] The global asymmetry analysis is close to tautological for the baselines. With shared/symmetric encoders and one embedding vector per node, swapping the node's role in (u,v) versus (v,u) leaves the representation computation unchanged by construction, so an asymmetry score of 0.00 is a mathematical identity, not an empirical demonstration of a representational bottleneck. DyGnROLE's nonzero scores are also expected because role embeddings are explicitly added. This analysis should be removed or reframed as a sanity check; the structural probing in Figure 3 is the more informative experiment. If Table 2 is kept, add variance and a control with RSPE removed.
minor comments (4)
  1. [Abstract / §3.3] The abstract names the pretraining objective DRA, while the body and §3.3 call it TCLP. Please standardize the terminology.
  2. [Table 1] No standard deviations are reported despite the statement that metrics are averaged over five seeds. Include them or confidence intervals, as several margins are small.
  3. [§4.2] The paper calls the 10k-label regime 'low-resource' but does not compare to other label budgets. A label-efficiency curve would substantiate the limited-label claim and clarify where pretraining helps.
  4. [Appendix A] Please state whether the same hyperparameters and tuning procedure were used for all baselines, and specify the baseline implementations (e.g., via DyGLIB) so that the comparisons are reproducible.

Circularity Check

1 steps flagged

Core SOTA claim is empirical and not circular; the only definitional step is the global asymmetry analysis, which restates the architecture by construction and is acknowledged as such.

specific steps
  1. self definitional [§5.3, Table 2 (Global Asymmetry)]
    "We calculate the score by setting h1 to the embedding of node u acting as the source in interaction (u, v), and h2 to the embedding of the same node u acting as the destination in the inverse interaction (v, u). A score of 0 implies perfect symmetry... As shown in Table 2, all baselines exhibit asymmetry scores of 0.00... Because this shared parameter space maps both source and destination queries through the exact same weights, the transformation is identical regardless of the node's functional role. Consequently, the model generates perfectly symmetric representations."

    For any baseline that maintains one node state and applies identical weights to source and destination queries, h1 and h2 are the same vector, so 1 - cos(h1,h2) = 0 exactly. The Table 2 baseline result is therefore entailed by the definition of the score and by the shared-parameter architecture; it is not an independently measured discovery of role blindness. Similarly, DyGnROLE's nonzero score is built in by its separate embedding tables and role-semantic positional encodings. The analysis restates the architectural difference rather than providing independent confirmation. The paper itself calls it a 'direct mathematical consequence,' so the step is acknowledged and is not load-bearing for the main SOTA claim.

full rationale

The paper's central claim is an empirical benchmark result, not a derivation. TCLP is a standard InfoNCE-style contrastive objective run on unlabeled historical edges, and the downstream edge-classification labels are held out from pretraining; no fitted parameter is renamed as a prediction. There is no load-bearing self-citation chain or imported uniqueness theorem. The main comparison does have a genuine protocol confound: DyGnROLE receives full unlabeled interaction history via TCLP pretraining while the baselines are trained from scratch on the 10k labeled edges, and the ablation shows pretraining is the dominant factor. That is a fairness/validity concern about attributing the gains to the role-aware architecture, not a circularity of the derivation. The only definitional step is the Global Asymmetry analysis in §5.3, where the baselines' 0.00 scores follow by construction from their shared-parameter design and from the score definition; the paper explicitly acknowledges this. This is a minor, non-central tautology. Overall, no significant circularity in the main empirical claim.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new natural-kind entities are postulated; all components are architectural modules validated by ablations. The empirical claim rests on the DTGB benchmark protocol, the masking assumptions of TCLP, and the anti-leakage masking of target edge features. TCLP is a standard InfoNCE objective and is not circular in the derivation sense.

free parameters (4)
  • Neighbor sample size k = 10
    Sets the input sequence length for source/destination history; fixed across all datasets; chosen by hand (§3.2, Appendix A).
  • Neighbor frequency count vocabulary threshold N_min = 10,000
    Only count values occurring ≥10,000 times in the training set get distinct indices; rarer counts map to UNK (§3.2). This data-dependent cutoff is a hand-chosen hyperparameter.
  • TCLP temperature tau = 0.07
    Temperature in the InfoNCE objective (§3.3, Appendix A); a standard hand-picked value.
  • Global hyperparameter configuration = 2 layers, 2 heads, d_c=50, d_t=100, dropout 0.1, lr 1e-4, batch 256
    One configuration is used for all eight datasets (Appendix A); hyperparameters are not tuned per dataset, so any dataset-specific gain could be sensitive to this fixed choice.
axioms (5)
  • domain assumption The DTGB chronological splits and 10k-label finetuning protocol are a valid measure of future edge classification performance.
    The entire evaluation rests on the DTGB benchmark's splits and protocol (§3.1, §4.2). If these splits leak or are misaligned, all reported scores are affected.
  • ad hoc to paper Historical neighbors are semantically valid relationships and should be masked, not treated as negatives, in contrastive pretraining.
    TCLP's historical false-negative masking assumes that any destination node previously connected to a source is a valid positive pair (§3.3). If historical interactions include spurious or negative links, the pretraining geometry would be distorted.
  • domain assumption The target edge's attributes are fully suppressed at the prepended position and no leakage occurs through shared node/edge text.
    In Feature Construction, the edge at index 0 is zero-masked to prevent label leakage (§3.2). The validity of edge classification depends on this masking being watertight.
  • domain assumption Static TinyBERT embeddings of node/edge text provide sufficient input signal.
    Node and edge texts are initialized once with TinyBERT and not updated (§3.2). The method assumes frozen text features are adequate.
  • standard math The Transformer, Fourier time encoding, and InfoNCE are taken as correct building blocks.
    Standard components from prior literature (§3.2, §3.3).

pith-pipeline@v1.3.0-alltime-deepseek · 11491 in / 16380 out tokens · 145148 ms · 2026-08-02T20:26:22.905558+00:00 · methodology

0 comments
read the original abstract

Edge classification on directed dynamic graphs requires modeling interactions between source and destination nodes exhibiting asymmetrical behavioral patterns and temporal dynamics. However, existing dynamic graph architectures largely rely on shared parameters for processing source and destination nodes, with limited or no systematic role-aware modeling. We propose DyGnROLE (Dynamic Graph Node-Role-Oriented Latent Encoding), a Transformer-based architecture that disentangles source and destination representations. By using separate embedding tables and role-semantic positional encodings, the model captures the distinct structural and temporal contexts unique to each role. Critical in limited-label settings, which are common in edge classification, is a self-supervised pretraining objective we introduce: Directional Role Alignment (DRA). DRA learns distinct but aligned source and destination embedding spaces by training source representations to retrieve their corresponding destination representations while a historical positive masking strategy excludes previously observed interactions from future negative comparisons. The masks introduce a temporally directional training signal in which node pairs progress monotonically from unseen to observed, after which the relationship is eligible only for further alignment. A comprehensive evaluation on four edge classification tasks across eight datasets demonstrates that DyGnROLE consistently outperforms a wide range of state-of-the-art baselines, highlighting the importance of role-aware representation learning and asymmetric pretraining for modeling complex directed interactions when labeled data is limited.

Figures

Figures reproduced from arXiv: 2602.23135 by Marek Rei, Tyler Bonnet.

Figure 1
Figure 1. Figure 1: Overview of the DyGnROLE architecture. The model begins processing a query interaction between a source (SRC) and destination (DST) node by retrieving their k most recent historical one-hop neighbors and transforming them into multi-faceted feature sets including node text (N), edge text (E), timestamps (T), within-sequence frequency (W), and cross-sequence frequency (C). These features are then linearly p… view at source ↗
Figure 2
Figure 2. Figure 2: Ablation study results using the remove-one strategy across eight datasets. We compare the performance of the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Asymmetry Scores (1−CosineSim) probing DyGFormer’s neighbor co-occurrence encoder and DyGnROLE’s neighbor fre￾quency embeddings. swapping the input nodes (u, v) to (v, u) and comparing the resulting feature vectors of the nodes in the first position. Results shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Label distributions of DTGB datasets [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 6 linked inside Pith

  1. [1]

    Dvgmae: Self- supervised dynamic variational graph masked autoen- coder.IEEE Transactions on Neural Networks and Learn- ing Systems,

    [Gaoet al., 2025 ] Mengzhou Gao, Xinxun Zhang, Pengfei Jiao, Tianpeng Li, and Zhidong Zhao. Dvgmae: Self- supervised dynamic variational graph masked autoen- coder.IEEE Transactions on Neural Networks and Learn- ing Systems,

    2025

  2. [6]

    Representation learning with contrastive predic- tive coding.arXiv preprint arXiv:1807.03748,

    [Oordet al., 2018 ] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic- tive coding.arXiv preprint arXiv:1807.03748,

    Pith/arXiv arXiv 2018

  3. [8]

    Towards better evaluation for dynamic link prediction

    [Poursafaeiet al., 2022 ] Farzan Poursafaei, Shenyang Huang, Kevin Pelrine, and Reihaneh Rabbany. Towards better evaluation for dynamic link prediction. InAdvances in Neural Information Processing Systems, volume 35, pages 32928–32941,

    2022

  4. [10]

    Dysat: Deep neural rep- resentation learning on dynamic graphs via self-attention networks

    [Sankaret al., 2020 ] Aravind Sankar, Yanhong Wu, Liang Gou, Wei Zhang, and Hao Yang. Dysat: Deep neural rep- resentation learning on dynamic graphs via self-attention networks. InProceedings of the 13th International Con- ference on Web Search and Data Mining (WSDM), pages 519–527. ACM,

    2020

  5. [11]

    Dyrep: Learn- ing representations over dynamic graphs

    [Trivediet al., 2019 ] Rakshit Trivedi, Mehrdad Farajtabar, Prasenjeet Biswal, and Hongyuan Zha. Dyrep: Learn- ing representations over dynamic graphs. InInternational Conference on Learning Representations (ICLR),

    2019

  6. [12]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    [Vaswaniet al., 2017 ] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Sys- tems, volume 30,

    2017

  7. [15]

    Inductive repre- sentation learning on temporal graphs.arXiv preprint arXiv:2002.07962,

    [Xuet al., 2020 ] Da Xu, Chao Ruan, Evren Korpeoglu, Srijan Kumar, and Kannan Achan. Inductive repre- sentation learning on temporal graphs.arXiv preprint arXiv:2002.07962,

    Pith/arXiv arXiv 2020

  8. [16]

    Cldg: Con- trastive learning on dynamic graphs

    [Xuet al., 2023 ] Yifei Xu, Bowen Shi, Tengfei Ma, Bo Dong, Hongyu Zhou, and Qing Zheng. Cldg: Con- trastive learning on dynamic graphs. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 696–707. IEEE, April

    2023

  9. [18]

    Dtgb: A comprehensive benchmark for dynamic text-attributed graphs

    [Zhanget al., 2024 ] Jiasheng Zhang, Jialin Chen, Menglin Yang, Aosong Feng, Shuang Liang, Jie Shao, and Rex Ying. Dtgb: A comprehensive benchmark for dynamic text-attributed graphs. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 91405– 91429. Curran Associates, Inc.,

    2024

  10. [20]

    The underlying raw data originates from publicly available sources, cited below

    A DyGnROLE Hyperparameters • Channel embedding dimension (d c): 50 • Temporal feature dimension (d t): 100 • Number of Transformer layers: 2 • Number of attention heads: 2 • Dropout rate: 0.1 • Maximum input sequence length: 10 • TCLP temperature 0.07 • Optimizer: AdamW • Weight decay: 0.01 • Learning rate: 1e-4 • Batch size: 256 B Dataset Details This se...

    2024

  11. [2002]

    GDELT4 is built from the Global Database of Events, Lan- guage, and Tone, recording international political events

    •Nodes: Employees •Node texts: Email addresses •Edges: Email communication between employees •Edge texts: Email contents •Edge labels: 10 communication types (e.g., ’calendar’, ’deal’). GDELT4 is built from the Global Database of Events, Lan- guage, and Tone, recording international political events. •Nodes: Political entities •Node texts: Names •Edges: A...

    2018

  12. [2016]

    Tinybert: Distilling bert for natural language under- standing

    [Jiaoet al., 2020 ] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language under- standing. InFindings of the Association for Computational Linguistics: EMNLP 2020, pages 4163–4174, November

    2020

  13. [2017]

    Tcl: Transformer-based dynamic graph modelling via contrastive learning.arXiv preprint arXiv:2105.07944,

    [Wanget al., 2021a ] Lu Wang, Xiaofu Chang, Shuang Li, Yunfei Chu, Hui Li, Wei Zhang, Xiaofeng He, Le Song, Jingren Zhou, and Hongxia Yang. Tcl: Transformer-based dynamic graph modelling via contrastive learning.arXiv preprint arXiv:2105.07944,

    Pith/arXiv arXiv

  14. [2018]

    Schardl, and Charles E

    [Parejaet al., 2020 ] Aldo Pareja, Giacomo Domeniconi, Jie Chen, Tengfei Ma, Toyotaro Suzumura, Hiroki Kaneza- shi, Tim Kaler, Tao B. Schardl, and Charles E. Leiser- son. Evolvegcn: Evolving graph convolutional networks for dynamic graphs. InProceedings of the AAAI Confer- ence on Artificial Intelligence, volume 34, pages 5363–

    2020

  15. [2019]

    [Liuet al., 2025 ] Weixiong Liu, Junwei Cheng, Quanlong Guan, Zhongyu Pan, and Chaobo He

    ACM. [Liuet al., 2025 ] Weixiong Liu, Junwei Cheng, Quanlong Guan, Zhongyu Pan, and Chaobo He. Dygmae: A novel dynamic graph masked autoencoder for link prediction. In UAI,

    2025

  16. [2020]

    Predicting dynamic embedding trajectory in temporal interaction networks

    [Kumaret al., 2019 ] Srijan Kumar, Xikun Zhang, and Jure Leskovec. Predicting dynamic embedding trajectory in temporal interaction networks. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1269–1278, Anchorage, AK, USA, July

    2019

  17. [2021]

    Inductive representation learn- ing in temporal networks via causal anonymous walks

    [Wanget al., 2021b ] Yu Wang, Yi-Ying Chang, Yang Liu, Jure Leskovec, and Pan Li. Inductive representation learn- ing in temporal networks via causal anonymous walks. arXiv preprint arXiv:2101.05974,

    Pith/arXiv arXiv

  18. [2022]

    Chambers, Fabrizio Frasca, and et al

    [Rossiet al., 2020 ] Emanuele Rossi, Benjamin P. Chambers, Fabrizio Frasca, and et al. Temporal graph networks for deep learning on dynamic graphs.arXiv preprint arXiv:2006.10637,

    Pith/arXiv arXiv 2020

  19. [2023]

    Towards better dynamic graph learning: New archi- tecture and unified library

    [Yuet al., 2023 ] Le Yu, Leilei Sun, Bowen Du, and Weifeng Lv. Towards better dynamic graph learning: New archi- tecture and unified library. InAdvances in Neural Infor- mation Processing Systems (NeurIPS), volume 36, pages 67686–67700,

    2023

  20. [2024]

    Topology-monitorable contrastive learning on dynamic graphs

    [Zhuet al., 2024 ] Zhenyu Zhu, Kunpeng Wang, Huan Liu, Jia Li, and Shijun Luo. Topology-monitorable contrastive learning on dynamic graphs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4700–4711. ACM, August

    2024

  21. [2025]

    Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415,

    [Hendrycks, 2016] Dan Hendrycks. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415,

    Pith/arXiv arXiv 2016