acceptodds
Under review as a conference paper at ICLR 2027

Transfer Learning for Edge Classification on Dynamic Text-Attributed Graphs

Abstract

Learning transferable representations for dynamic text-attributed graphs (DyTAGs) requires models to capture underlying interaction dynamics that persist across domains. However, existing methods tend to overfit to domain-specific structural, temporal, and semantic patterns, limiting edge classification performance under distribution shifts. To expose and address this, we formally establish a leave-one-domain-out (LODO) transfer learning protocol for edge classification on DyTAGs. Under this protocol, we demonstrate that state-of-the-art self-supervised methods for dynamic graph learning perform poorly when transferred to unseen domains. Strikingly, existing methods underperform a structurally and temporally unaware Bag of Events (BoE) model we introduce, which inputs only unordered sequences of node and edge text features. Proceeding from the BoE, we propose Spatio-Temporal Semantic Alignment (STSA), which integrates a spatio-temporal encoder that fuses representations of time deltas and node occurrence frequencies into a unified manifold. STSA is trained with a Contrastive Semantic Forecasting objective, which anchors edge representations to a multi-domain textual latent space initialized by a pretrained language model, providing a robust prior that outperforms BoE and all existing methods we evaluate.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic Graph Representation Learning, Text-Attributed Graphs, Transfer Learning, Self-Supervised Learning, Edge Classification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gEWDOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates cross-domain edge classification on dynamic text-attributed graphs. It introduces a domain-wise holdout self-supervised pre-training pipeline for few-shot adaptation on target domains, with Bag of Events (BoE) designed as a baseline without explicit temporal-spatial encoding. The proposed STSA integrates time offsets and node occurrence frequencies via MLP, and leverages Contrastive Semantic Forecasting (CSF) to predict interaction text embeddings. The topic is meaningful. The experimental setup, covering four domains, eight datasets and frozen-encoder evaluation, forms a solid foundation.

Critical strengths
  1. The research question carries practical value. By combining cross-domain unsupervised pre-training with target-domain edge classification, this work can examine whether the model learns reusable interaction representations, rather than merely fitting prediction patterns on a single graph.

  2. The BoE model achieves strong performance even with weakened explicit spatiotemporal encoding, demonstrating that complex structural designs do not guarantee improved transferability. This finding offers valuable insights to re-examine the respective roles of semantics, graph structure, and temporal information in dynamic graph models.

Critical weaknesses
  1. The paper attributes the weaker performance of complex methods compared to BoE to overfitting on the source domain’s structural, temporal and linguistic distributions. However, Table 1 shows that most baselines already approach majority-class prediction without pre-training, and pre-training yields only marginal or no improvement. This suggests an alternative explanation: the target-domain training setup, optimization conditions or input adaptation prevent these methods from learning effective classifiers. Their failure to generalize does not emerge after pre-training.

  2. The number of classes across the eight tasks varies widely from 2 to 266. Fixing the label pool to the most recent 500 instances creates substantially different learning setups. For binary tasks, this may yield sufficient class coverage. However, 147 classes in GDELT and 169 classes in ICEWS are absent from the fine-tuning samples. In addition, using consecutive recent samples can introduce strong temporal locality bias.

Questions for authors

See Weakness

1 comment

We thank Reviewer gEWD for their time carefully reviewing our work and providing useful feedback. Your primary concerns overlap with those raised by fellow reviewers, and are therefore addressed in the General Response at the top to avoid repetition.

Mapping of Shared Concerns to the General Response:

  • 1. Majority Prediction and Source-Domain Overfitting: "Table 1 shows that most baselines already approach majority-class prediction without pre-training... This suggests an alternative explanation: the target-domain training setup, optimization conditions or input adaptation prevent these methods from learning effective classifiers."
  • 4. Label Budget, Unseen Classes, and Temporal Locality: "147 classes in GDELT and 169 classes in ICEWS are absent from the fine-tuning samples. In addition, using consecutive recent samples can introduce strong temporal locality bias."

Because both of your specific weaknesses (and corresponding questions) are fully covered by the new control experiments and discussion in these two sections of the General Response, we kindly refer you there for our complete response.

Reviewer_f1oROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies transfer learning for edge classification on dynamic text-attributed graphs. It introduces a leave-one-domain-out protocol, a Bag of Events (BoE) baseline, and STSA, which combines spatio-temporal features with contrastive prediction of query-edge text embeddings. The evaluation covers eight datasets using finetuning and frozen-encoder linear probing.

Critical strengths

This paper provides a useful setting for cross-domain edge classification and reports consistent gains across eight datasets.

  • The leave-one-domain-out protocol tests transfer to unseen domains. BoE provides a simple reference for assessing more complex graph models.

  • STSA has the best reported Macro F1 on all eight datasets, averaging 24.22% versus 22.93% for BoE. The results include ten runs and frozen-encoder K-shot evaluation.

Critical weaknesses

Motivation and design

  • Table 1 lacks target-only pretraining. With unlabeled target history available under Eq. (1), what does cross-domain pretraining add? Compare using history before the labeled adaptation window, with matched encoders, objectives, and compute.

  • Section 4.3 blames source-domain overfitting, yet GCAL and IDOL already match the majority baseline without pretraining on five datasets. Most methods improve after pretraining. Why is overfitting considered the cause?

  • BoE selects one-hop interactions before the query time (Section 3.3). Its inputs still use topology and history. Its results do not show that structural or temporal information hurts transfer.

Experiments

  • Removing STE lowers Macro F1 from 24.22 to 22.81 in Table 2, but retains time and frequency inputs. Please compare with BoE+CSF using joint history processing and matched capacity to assess what the spatio-temporal inputs add.

  • The experiments use only Qwen3-Embedding-4B (released June 5, 2025, over 16 months ago). Do STSA's gains hold with a larger encoder such as Qwen3-Embedding-8B or a different encoder such as pplx-embed-v1-4b? Compare BoE, BoE+CSF, and STSA with the same encoder, updating inputs and any CSF targets. Keep history, query-edge masking, label budget, and dimensions fixed. Report accuracy and embedding cost.

  • Section 4.2 uses 500 labels and 20 finetuning epochs. Were baseline learning rates and training length tuned on validation data? Appendix G does not report this search.

  • Figure 2 measures rank before the Transformer on four 60-node graphs. Does the difference remain after the Transformer on real target graphs? Higher rank alone does not explain better transfer. Compare with separate time and frequency MLPs at the same parameter budget.

Questions for authors
  • How does source pretraining compare with target-only pretraining on earlier history at comparable compute?

  • How does BoE perform with CSF, joint history processing, and comparable capacity?

  • Do STSA's gains over BoE and BoE+CSF remain with Qwen3-Embedding-8B or pplx-embed-v1-4b? Please update inputs and any CSF targets, keep the query edge masked, and report accuracy and embedding cost.

  • Were learning rates and training length tuned on validation data for the baselines at majority-level performance?

1 comment

We thank Reviewer f1oR for their thoughtful review and helpful feedback. Several concerns were shared by fellow reviewers, and are therefore addressed in the General Response at the top to avoid repetition. We address your remaining questions regarding the BoE baseline and text encoder generalization below.

Mapping of Shared Concerns to the General Response:

  • 1. Majority Prediction and Source-Domain Overfitting: "Section 4.3 blames source-domain overfitting, yet GCAL and IDOL already match the majority baseline without pretraining..."
  • 2. Target-Only Pretraining Baseline: "Table 1 lacks target-only pretraining..."
  • 3. Model selection: "Were baseline learning rates and training length tuned on validation data?"
  • 5. Additional Capacity-Matched Ablations: "Please compare with BoE+CSF... Compare with separate time and frequency MLPs..."
  • 6. Real-Dataset Representational Analysis: "Does the difference remain after the Transformer on real target graphs?"

Motivation 3: Use of Topology and History in BoE.
"BoE selects one-hop interactions before the query time (Section 3.3). Its inputs still use topology and history. Its results do not show that structural or temporal information hurts transfer."

We agree the results of BoE do not show structural or temporal information hurts transfer. We do not claim in the paper that such information is inherently harmful. In fact, our proposed STSA framework relies on spatio-temporal features. Regarding the use of topology in BoE, we would point out that while BoE does retrieve past interactions, it does not need a graph representation to do so. It simply processes a flat, unordered sequence of recent events for the queried entities. It does not process the structural properties typically associated with graph topology. Ultimately, the purpose of BoE as a control is not to prove that structural and temporal signals hurt transfer, but rather to demonstrate that the complex parameterizations of these signals in existing baselines fail to provide an advantage over this simple baseline.

We have clarified this framing in the revised manuscript (Section 1, lines 56–59; Section 3.3, lines 138–140; Section 4.3, lines 400–403).

Experiments 2 & Q3: Text Encoding Generalization.
"Do STSA's gains over BoE and BoE+CSF remain with Qwen3-Embedding-8B or pplx-embed-v1-4b? Please update inputs and any CSF targets, keep the query edge masked, and report accuracy and embedding cost."

To verify whether the gains achieved by STSA over BoE and BoE + CSF hold across different text encoders, we evaluated these models using texts embedded by pplx-embed-v1-4b. The table below reports average Macro-F1 and Accuracy for these runs. As requested, we include Accuracy alongside our primary metric, Macro-F1. Because many DTGB datasets suffer from severe class imbalance, Accuracy disproportionately rewards models that tend to over-predict the majority class (please see Appendix E for DTGB edge class distributions). Consequently, BoE achieves higher average Accuracy despite its lower Macro-F1. The results show that the relative Macro-F1 performance gaps between BoE, BoE + CSF, and STSA remain consistent across both text encoders. Regarding cost, for both encoders, embedding all unique texts for interactions used in the experiments takes approximately 48.2 GPU-hours with 9.7 GB peak VRAM. Please see Appendix H for details on computing hardware.

Average performance comparison across all DTGB datasets using Qwen3-Embedding-4B and pplx-embed-v1-4b.

Text EncoderMethodMacro-F1 (%)Accuracy (%)
Qwen3-Embedding-4BBoE22.93$_{\pm 0.94}$47.44±1.17_{\pm 1.17}
BoE + CSF23.86$_{\pm 0.40}$47.06$_{\pm 0.89}$
STSA24.22±0.44_{\pm 0.44}47.18$_{\pm 0.88}$
pplx-embed-v1-4bBoE22.39$_{\pm 0.86}$47.68±1.21_{\pm 1.21}
BoE + CSF23.26$_{\pm 0.36}$46.39$_{\pm 0.77}$
STSA23.43±0.41_{\pm 0.41}46.79$_{\pm 0.70}$

We include the Macro-F1 comparison across text encoders in Appendix P of the revised manuscript.

Reviewer_9r9wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies cross-domain transfer for edge classification on dynamic text-attributed graphs, using a leave-one-domain-out protocol that pretrains on source domains and then transfers to a held-out one with few labels. The authors introduce a strong text-only baseline and a model that adds spatio-temporal and semantic components on top of it. On eight datasets the proposed model has the best average score.

Critical strengths

S1. The protocol and the text-only baseline address a useful gap — prior work on transfer in dynamic graphs has addressed link prediction and node- or graph-level tasks, but not edge classification. A text-only baseline beats several specialized dynamic graph methods, which is surprising.

S2. The proposed model improves on the text-only baseline consistently, with significance over the strongest baseline on seven of the eight datasets. The reported statistical tests and reference points are appreciated.

Critical weaknesses

W1. The paper blames the baselines' poor performance on source-domain overfitting, but Table 1 doesn't support it. GCAL, IDOL and DyGMAE score the same as Majority on five datasets, with and without pretraining, and TLP does too, apart from 0.01 on Google with pretraining. Random averages 18.26, above every existing method except pretrained CrossLink and DyGnROLE. They collapse even without pretraining, so overfitting isn't the reason. The comparison is also under-specified: it's unclear whether the baselines get the same text features, and nothing is said about the finetuning learning rate, class-imbalance handling, model selection or tuning ranges. An in-domain pretraining reference would separate domain shift from the difficulty of learning with few target labels. The fixed 500-label budget also leaves 147 of 236 GDELT classes and 169 of 266 ICEWS classes unseen during finetuning.

W2. The ablations show that the spatio-temporal encoder and the contrastive objective help, but they don't explain where the gain comes from. The eRank analysis only shows that the encoder changes the input geometry, not that this geometry rather than the extra nonlinear layer is what improves transfer, and eRank is never measured on the real datasets. The Enron comparison changes several things at once — the text-only baseline drops the time and frequency inputs along with the encoder. Removing the encoder (22.81) or swapping CSF for InfoNCE (22.99) lands close to the baseline (22.93), which is consistent with the two components being complementary, but doesn't explain where the gain comes from. A text-only baseline with the contrastive objective, and a capacity-matched model that handles time and frequency separately, would separate the two. Otherwise the paper should narrow the claim to what the ablations support.

Questions for authors

Q1. Do all baselines use the same node and historical edge text representations as the proposed model and its text-only baseline? If any baseline uses node-ID embeddings or memory states, how are these initialized or adapted for the target domain? Please also give the finetuning learning rate, class-imbalance handling, model selection and tuning ranges. Appendix E lists average rating and review count in Yelp user texts and rank score in Amazon node texts — were these fields available at each query time, or taken from a full snapshot?

Q2. Could you provide an in-domain pretraining reference for the proposed model, its text-only baseline and the strongest baseline on one representative target domain? This would test the source-overfitting explanation; if it is not supported, narrowing that conclusion would be a reasonable response.

Q3. Could you report two additional runs — the text-only baseline with the contrastive objective, and a capacity-matched model that processes time and frequency separately? Also, does the CSF to InfoNCE ablation process the two sequences jointly or independently, and if jointly, how does it avoid the cross-attention shortcut described in Section 3.3?

Q4. What do the Seen and Unseen bars in Figure 7 measure, and what are their numerical values, evaluation subsets and label sets? If the hatched bars show the full-test Macro-F1 from Table 1, please explain the differences. Do the ten runs repeat both pretraining and finetuning, or only finetuning?

3 comments

We thank Reviewer 9r9w for their detailed review and constructive feedback. Several concerns were shared by fellow reviewers, and are therefore addressed in the General Response at the top to avoid repetition. We address your remaining specific questions regarding input parity, snapshot fields, and sequence processing below.

Mapping of Shared Concerns to the General Response:

  • 1. Majority Prediction and Source-Domain Overfitting: "The paper blames the baselines' poor performance on source-domain overfitting, but Table 1 doesn't support it..."
  • 2. Target-Only Pretraining Baseline: "Could you provide an in-domain pretraining reference...?"
  • 3. Model Selection: "...nothing is said about the finetuning learning rate, class-imbalance handling, model selection or tuning ranges."
  • 4. Label Budget, Unseen Classes, and Temporal Locality: "The fixed 500-label budget also leaves 147 of 236 GDELT classes and 169 of 266 ICEWS classes unseen during finetuning."
  • 5. Additional Capacity-Matched Ablations: "A text-only baseline with the contrastive objective, and a capacity-matched model that handles time and frequency separately, would separate the two..."
  • 6. Real-Dataset Representational Analysis: "eRank is never measured on the real datasets."

Q1: Input Parity, Inductive Adaptation, and Snapshot Fields

"Do all baselines use the same node and historical edge text representations as the proposed model and its text-only baseline? If any baseline uses node-ID embeddings or memory states, how are these initialized or adapted for the target domain? [...] Appendix E lists average rating and review count in Yelp user texts and rank score in Amazon node texts: were these fields available at each query time, or taken from a full snapshot?"

Input Parity. All evaluated models receive L2L_2-normalized 256-d Qwen3-Embedding-4B representations. STSA, BoE, and DyGnROLE process node and edge text embeddings. By their original design, GCAL, IDOL, DyGMAE, TLP, DVGMAE, CLDG, and EvoluNet do not process edge text features, while CrossLink zeroed both node and edge features due to heterogeneous raw dataset dimensions. All eight of these baselines receive the same 256-d Qwen3 node embeddings in our experiments, with CrossLink retaining zeroed edge features per its §C.1.

Inductive Adaptation. Three baselines originally used transductive node ID embeddings or persistent recurrent memory states, which cannot transfer to disjoint target nodes. For inductive evaluation, we adapted them as follows:

  • TLP: The original TLP maintains node-ID-indexed memory states in a per-node memory store (N×dMN \times d_M) updated recurrently by a GRU within its Temporal Graph Network (TGN) backbone. We removed this transductive memory store and GRU updater, replacing them with inductive snapshot GCN layers that compute node representations from the 256-d Qwen3 node text features and the causal snapshot adjacency. We retained its core structural mapping module, detailed in Sections 4.1–4.2 of the TLP paper, aligning node structural features (degree, centrality, and clustering) with the GCN node representations.
  • DyGMAE: The original DyGMAE maintains a recurrent hidden state with a GRU across all sequential snapshots of a graph, as defined in Section 3.3 of the DyGMAE paper. Because in our setup each minibatch contains interactions from a single source dataset (Appendix F) and persistent node-indexed states cannot transfer across domains, we instead run the GRU over a local 5-timestep causal snapshot window per batch, starting from a zero-initialized hidden state and discarding it afterward. The GRU still models temporal evolution across those 5 historical snapshots, while remaining stateless across batches and domains.
  • DVGMAE: The original DVGMAE also maintains a recurrent hidden state with a GraphGRU across snapshots, detailed in Section IV-C of the DVGMAE paper. We similarly zero-initialize the hidden state for each batch, running it over a 5-timestep causal snapshot window during pretraining and a single causal snapshot at inference. Additionally, when a graph has no node features, DVGMAE originally uses a one-hot encoding and learnable matrix (a node ID embedding) as described in Section IV-B. Because our datasets provide 256-d Qwen3 node features, this node-ID matrix is not used.

Snapshot Fields. The reviewer correctly notes that the rank scores in Amazon and the aggregate user statistics in Yelp (average ratings, number of reviews, and number of fans) were derived from full static snapshots in the original datasets rather than query-time states. To address the resulting risk of future-information leakage, we conducted a new control experiment where we removed these metadata fields, re-encoded the texts, and re-evaluated the models. As the table below demonstrates, removing this metadata causes a performance decrease for STSA and BoE on the Yelp dataset, but the overall model rankings and relative performance gaps remain consistent.

Clean-node-text control on Amazon and Yelp (Macro-F1 %), removing rank from Amazon product texts and average ratings, number of reviews, and number of fans from Yelp user texts.

Dataset / ConditionSTSABoEDyGnROLECrossLinkGCALIDOL
Amazon (Original)25.22±0.40_{\pm 0.40}21.87$_{\pm 1.71}$17.06$_{\pm 0.30}$15.76$_{\pm 0.00}$15.76$_{\pm 0.00}$15.76$_{\pm 0.00}$
Amazon (Cleaned)25.22±0.46_{\pm 0.46}21.84$_{\pm 1.77}$17.06$_{\pm 0.28}$15.76$_{\pm 0.00}$15.76$_{\pm 0.00}$15.76$_{\pm 0.00}$
Yelp (Original)28.96±0.75_{\pm 0.75}23.06$_{\pm 2.04}$16.94$_{\pm 1.40}$18.47$_{\pm 1.43}$13.99$_{\pm 0.00}$13.99$_{\pm 0.00}$
Yelp (Cleaned)27.59±0.68_{\pm 0.68}22.52$_{\pm 2.32}$17.07$_{\pm 1.41}$18.42$_{\pm 1.47}$13.99$_{\pm 0.00}$13.99$_{\pm 0.00}$

In the revised manuscript, we include the input parity details and inductive adaptations in Appendix L (referenced in Section 4.2, lines 382–383), clarify the static snapshot node fields in Appendix E, and add the clean-node-text control table in Appendix K.

Q3: Sequence Processing in InfoNCE Ablation
"Does the CSF to InfoNCE ablation process the two sequences jointly or independently, and if jointly, how does it avoid the cross-attention shortcut described in Section 3.3?"

For the CSF to InfoNCE ablation, the two sequences are processed independently to avoid the cross-attention shortcut.

We have added this clarification in Section 4.5 (lines 465–466) of the revised manuscript.

Q4: Ten Runs and Variance
"Do the ten runs repeat both pretraining and finetuning, or only finetuning?"

For each model, we perform a single pretraining run and save the checkpoint, then perform ten independent finetuning runs loaded from that single pretraining checkpoint using different random seeds.

We have added this clarification in Section 4.2 (line 377) of the revised manuscript.

Comments on the paper

7 comments
AuthorsOpenReview ↗

6. Real-Dataset Representational Analysis

Addresses 9r9w (W2) and f1oR (Experiments 4)

To address concerns about the representational analysis (Section 4.1), we evaluated pre- and post-Transformer eRank on 10,000 real held-out target edges from all DTGB datasets.

The results in the table below show that the pre-Transformer eRank ordering observed on synthetic graphs holds on real DTGB datasets: STSA achieves the highest Pre-Tr eRank (46.26), followed by STSA (Non-Linear ST) (32.57), STSA (Linear ST) (26.88), DyGnROLE (20.36), and CrossLink (14.62). After passing through the Transformer layers, all three STSA variants reach substantially higher Post-Tr eRank (76.43–79.89) than DyGnROLE (56.18) and CrossLink (30.08). Among the STSA variants themselves, Post-Tr eRank converges to a similar range (76.43–79.89) once self-attention mixes the spatio-temporal and text features.

Reviewer f1oR correctly notes that higher rank alone does not explain better performance. We do not claim that models achieve better performance because of a higher eRank. Instead, eRank serves as a diagnostic metric to analyze the mechanistic behavior of the network, demonstrating how different encoding strategies shape the geometric representation space. While a higher eRank merely confirms that a wider variety of non-redundant features is being extracted, a concurrent increase in downstream Macro-F1 is necessary to confirm that those extracted features are predictively useful. Because many real-world dynamic graph evaluations lack this type of representational analysis, it is a valuable addition to the second contribution of the paper, and we thank the reviewer for the suggestion.

Real-dataset pre- and post-Transformer Effective Rank (eRank) of the spatio-temporal (time and node frequency) features across DTGB target graphs (10,000 held-out target edges per graph test split).

MethodPre-Tr eRankPost-Tr eRank
CrossLink14.6230.08
DyGnROLE20.3656.18
STSA (Linear ST)26.8879.89
STSA (Non-Linear ST)32.5776.43
STSA46.2677.39

We have updated Section 4.1 of the revised manuscript to clarify the diagnostic purpose of this analysis (lines 267–269) and report the real-dataset pre-Transformer eRank results (lines 321–324).

AuthorsOpenReview ↗

5. Additional Capacity-Matched Ablations

Addresses 9r9w (W2, Q3) and f1oR (Experiments 1, Q2, Experiments 4)

Reviewers questioned whether the performance gains of STSA were driven primarily by the pretraining objective (CSF), the non-linear capacity of the Spatio-Temporal Encoder (STE), or the cross-modal fusion of the STE. To isolate these mechanisms, we conducted three additional ablations.

First, we applied the CSF objective to the structurally and temporally unaware BoE baseline (BoE w/ CSF). While this configuration improves upon the standard BoE baseline (raising average Macro-F1 from 22.93% to 23.86%), it underperforms the variants with non-linear spatio-temporal encodings. This indicates that semantic alignment is most effective when grounded in non-linear spatio-temporal representations.

Second, we isolated the structural encoding mechanisms by replacing the non-linear STE with a linear projection of the features (STSA w/ Linear ST). The resulting performance drop indicates that non-linear feature projection is beneficial for capturing complex spatio-temporal dependencies.

Finally, to address the hypothesis of Reviewer 9r9w regarding cross-modal fusion, we replaced the STE with separate MLPs for time encodings and frequency embeddings, matching the parameter count of the STE (STSA w/ Non-Linear ST). We provide our updated ablation results table below. On average, the results support the reviewer's hypothesis: Complete STSA (24.22%) and STSA w/ Non-Linear ST (24.17%) perform comparably, indicating that non-linear capacity is the primary driver of the overall performance gain over BoE w/ CSF (23.86%). However, examining the per-dataset results reveals an important difference in consistency: on Stack U, both unfused variants, STSA w/ Linear ST (51.42%) and STSA w/ Non-Linear ST (51.47%), underperform BoE (51.61%), whereas Complete STSA outperforms BoE across all eight DTGB datasets. Accordingly, we retain Complete STSA as our proposed framework and update our claims in the revised manuscript: non-linear capacity drives the average performance gain, while pre-Transformer cross-modal fusion provides more consistent benefits.

Ablation study of STSA finetuning for dynamic edge classification. Values report average Macro F1 (%) ±\pm standard deviation for individual datasets over 10 runs with different random seeds. Best performance in each column is shown in bold.

VariantAmazonEnronGDELTGoogleICEWSStack EStack UYelpAvg
Reference Baseline
BoE24.61$_{\pm 0.86}$7.71$_{\pm 0.38}$0.28$_{\pm 0.03}$18.11$_{\pm 1.16}$0.54$_{\pm 0.11}$54.22$_{\pm 1.34}$51.61$_{\pm 1.27}$26.34$_{\pm 2.34}$22.93$_{\pm 0.94}$
BoE + CSF25.37$_{\pm 0.26}$8.18$_{\pm 0.33}$0.39$_{\pm 0.03}$20.15$_{\pm 1.04}$1.10$_{\pm 0.07}$54.16$_{\pm 0.59}$52.42$_{\pm 0.55}$29.12$_{\pm 0.34}$23.86$_{\pm 0.40}$
STSA Variants
CSF →\to SF23.79$_{\pm 0.47}$8.34$_{\pm 0.42}$0.41$_{\pm 0.02}$20.23$_{\pm 0.72}$1.12$_{\pm 0.06}$54.38$_{\pm 2.29}$52.71$_{\pm 0.78}$28.16$_{\pm 0.38}$23.64$_{\pm 0.64}$
CSF →\to InfoNCE21.72$_{\pm 0.62}$10.53±1.21_{\pm 1.21}0.37$_{\pm 0.01}$18.47$_{\pm 0.79}$0.95$_{\pm 0.06}$54.08$_{\pm 1.22}$50.88$_{\pm 0.94}$26.97$_{\pm 0.54}$22.99$_{\pm 0.68}$
Linear ST24.30$_{\pm 0.43}$8.67$_{\pm 0.60}$0.44$_{\pm 0.01}$19.27$_{\pm 0.82}$1.22$_{\pm 0.05}$55.81$_{\pm 0.99}$51.42$_{\pm 0.96}$26.78$_{\pm 0.63}$23.49$_{\pm 0.56}$
Non-Linear ST25.49±0.57_{\pm 0.57}8.71$_{\pm 0.41}$0.44$_{\pm 0.02}$20.10$_{\pm 0.82}$1.20$_{\pm 0.07}$56.18±0.63_{\pm 0.63}51.47$_{\pm 0.82}$29.74±0.63_{\pm 0.63}24.17$_{\pm 0.50}$
Complete25.22$_{\pm 0.40}$9.12$_{\pm 0.23}$0.46±0.02_{\pm 0.02}20.40±0.72_{\pm 0.72}1.24±0.08_{\pm 0.08}55.52$_{\pm 0.55}$52.84±0.76_{\pm 0.76}28.96$_{\pm 0.75}$24.22±0.44_{\pm 0.44}

We include the updated ablation table (Table 2, lines 433–445) and results discussion (lines 457–463) in Section 4.5 of the revised manuscript.

AuthorsOpenReview ↗

4. Label Budget, Unseen Classes, and Temporal Locality

Addresses 9r9w (W1, Q4) and gEWD (W2)

It was noted that the 500-edge adaptation budget creates substantially different learning setups across datasets, particularly regarding GDELT and ICEWS, which have a large number of classes that are unseen in the 500 edges. There was also a concern that this short window of consecutive edges might introduce temporal locality bias.

To address the first concern, we ran finetuning on additional budgets up to 5000 and present the results for GDELT and ICEWS below, specifying the proportion of unseen classes at each budget.

Macro-F1 (%) on GDELT and ICEWS across label budgets (BB) from 500 to 5,000. # Test Classes Unseen reports how many classes present in the test split (220 of 236 total for GDELT; 236 of 266 total for ICEWS) are absent from the BB-edge finetuning window, while % Test Edges Unseen reports the percentage of test edges belonging to those unseen classes. Bold indicates the best result in each column.

DatasetMetric / ModelB=500B=500B=1,000B=1,000B=2,000B=2,000B=5,000B=5,000
GDELT# Test Classes Unseen14712510581
% Test Edges Unseen6.72%4.04%2.23%1.09%
GCAL0.050.060.110.14
CrossLink0.090.100.200.19
DyGnROLE0.270.280.330.39
BoE0.280.440.490.58
STSA0.460.630.670.77
ICEWS# Test Classes Unseen16914612188
% Test Edges Unseen6.83%3.86%2.29%0.91%
GCAL0.100.110.120.12
CrossLink0.120.120.170.77
DyGnROLE0.430.611.001.54
BoE0.541.041.592.19
STSA1.241.682.162.49

We would like to offer clarification regarding the Seen and Unseen bars in Figure 7 in Appendix I. The solid bars represent performance exclusively on the classes present in the 500-edge finetuning window (the "seen" label set). The hatched bars represent performance exclusively on the classes missing from those 500 edges (the "unseen" label set). The Table 1 results are not the hatched bars, but the aggregate macro-average of both the seen and unseen groups combined.

To address the temporal locality concern, we ran an experiment to test whether our results suffer from bias, i.e., whether models only performed well because the 500 finetuning edges were taken from the very end of the training timeline close to the test set. To evaluate this, we shifted the contiguous 500-edge adaptation window across five chronological points in the training history ending at 20%, 40%, 60%, 80% of the original DTGB train split and tested on the DTGB test split. We present the results in the table below, including the results on the original 500-label window (ending at 100%) for easy reference. STSA consistently outperforms other evaluated models regardless of the temporal origin of the adaptation window.

Macro-F1 (%) using different 500-label windows for finetuning, sampled evenly across the DTGB train split ending at 20%, 40%, 60%, 80%, 100%. Bold indicates best in column, and Avg column includes Std across windows.

Method20%40%60%80%100%Avg ±\pm Std
GCAL16.9417.1217.1917.0517.1117.08$_{\pm 0.09}$
CrossLink18.4218.5818.1619.7118.6518.70$_{\pm 0.59}$
DyGnROLE18.9318.2319.2719.7219.2019.07$_{\pm 0.55}$
BoE20.8920.1121.2921.8422.9321.41$_{\pm 1.06}$
STSA23.2022.8023.5524.5324.2223.66$_{\pm 0.71}$

In the revised manuscript, we have clarified the Figure 7 caption in Appendix I, added the label budget scaling results in Appendix N (referenced in Section 4.3, lines 412–413), and added the temporal locality experiment in Appendix O (referenced in Section 4.2, line 376).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.