Research map / Graph learning / Knowledge and temporal graphs
Node-Time Conditional Prompt Learning in Dynamic Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DyGPrompt, a pre-train then prompt-learning framework for continuous-time dynamic graphs that uses dual node/time prompts and dual condition-nets to generate time-conditioned node prompts and node-conditioned time prompts for node classification and link prediction.
Problem
DGNNs are pre-trained on link prediction, creating a gap with downstream tasks like node classification, and existing graph prompt methods ignore temporal evolution and the mutual dependence between node and time patterns in dynamic graphs.
Approach
Dual prompts (a node feature prompt and a time feature prompt) bridge task and temporal gaps; dual condition-nets (a time condition-net producing time-conditioned node prompts and a node condition-net producing node-conditioned time prompts) model node-time interplay parameter-efficiently while the pre-trained DGNN stays frozen.
Key results
- TGN-DyGPrompt reaches 94.33% AUC on Wikipedia transductive link prediction versus 72.48% for TGN alone.
- TGAT-DyGPrompt achieves 82.09% AUC on Wikipedia node classification, above the best baseline TGAT-TIGPrompt at 69.21%.
- On DGraph anomaly detection, DyGPrompt reaches 73.39% AUC versus 64.42% for TIGPrompt.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DyGPrompt prompts and condition-nets to TGAT | TGAT · Wikipedia node classification | AUC-ROC (%): 67.00 -> 82.09 | Helped |
| add DyGPrompt prompts and condition-nets to TGN | TGN · Reddit transductive link prediction | AUC-ROC (%): 67.37 -> 96.82 | Helped |
| add DyGPrompt prompts and condition-nets to TGN | TGN · Wikipedia node classification | AUC-ROC (%): 50.61 -> 74.47 | Helped |
| add DyGPrompt to DGNN backbones | DyRep, JODIE, TGAT, TGN, TREND, and GraphMixer · node classification, transductive and inductive link prediction on Wikipedia, Reddit, and MOOC | AUC-ROC (%) | Mixed |
| add a node prompt (ablation) | TGAT Variant 1 · node classification on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 67.00 -> 72.59; 53.64 -> 61.82; 59.27 -> 63.50 | Helped |
| add a time prompt (ablation) | TGAT Variant 1 · node classification on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 67.00 -> 73.22; 53.64 -> 62.51; 59.27 -> 62.59 | Helped |
| add both node and time prompts without condition-nets (ablation) | TGAT Variant 1 · node classification on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 67.00 -> 72.25; 53.64 -> 63.11; 59.27 -> 62.87 | Helped |
| add a node condition-net (ablation) | TGAT Variant 2 with node prompt · node classification on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 72.59 -> 81.40; 61.82 -> 73.12; 63.50 -> 77.15 | Helped |
| add a time condition-net (ablation) | TGAT Variant 3 with time prompt · node classification on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 73.22 -> 80.34; 62.51 -> 72.59; 62.59 -> 76.16 | Helped |
| add both dual condition-nets to the dual-prompt variant (ablation) | TGAT Variant 4 · node classification on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 72.25 -> 82.09; 63.11 -> 73.50; 62.87 -> 77.78 | Helped |
| add DyGPrompt prompts and condition-nets (ablation) | TGAT Variant 1 · transductive link prediction on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 55.78 -> 69.88; 62.43 -> 90.76; 51.49 -> 53.92 | Helped |
| add DyGPrompt prompts and condition-nets (ablation) | TGAT Variant 1 · inductive link prediction on Wikipedia, Reddit, and MOOC | AUC-ROC (%): 48.21 -> 52.58; 57.30 -> 75.20; 51.42 -> 53.29 | Helped |
Surprising findings
- Conventional DGNNs TGAT and TGN score lower than originally reported under the limited-data downstream setting, which the authors attribute to fast convergence on 30 sampled events.
- On the MOOC dataset, differences between prompt variants are small, which the authors attribute to less dynamic interactions.
Limitations
- Evaluated only in data-scarce settings with 30 downstream events per task, so behavior with more labels is not shown.
- Gains are smaller on MOOC, where the authors suggest prompts are less effective for less dynamic interactions.
- Baseline DGNN results (TGAT, TGN) fall below originally reported numbers, so comparisons depend on the authors' custom split and protocol.
Experimental setup
Datasets: Wikipedia Reddit MOOC Genre DGraph ML-Rating
Benchmarks: Wikipedia Reddit MOOC Genre DGraph ML-Rating
Metrics: AUC-ROC
Baselines: ROLAND TGAT TGN TREND GraphMixer DDGCL CPDG GraphPrompt ProG TIGPrompt
Builds on: GraphPrompt TIGPrompt TGAT TGN conditional prompt learning
Closest papers by meaning
- One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 2025 · similarity 0.89)
- A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation (AAAI 2025 · similarity 0.88)
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.88)
- Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.88)
- Towards Dynamic Message Passing on Graphs (NeurIPS 2024 · similarity 0.88)
- GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.