Research map / Graph learning / Knowledge and temporal graphs
TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction
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 TAMI, a plug-in framework for temporal graph link prediction with a log time encoding (LTE) that rescales skewed time intervals and a link history aggregation (LHA) module that keeps the k most recent interactions per node pair, improving GraphMixer and DyGFormer.
Problem
Temporal graph link prediction degrades for node pairs with infrequent interactions because skewed interaction intervals hinder sinusoidal time encoding learning and recent-neighbor aggregation forgets pairwise history.
Approach
Apply a logarithmic transform ln(1+Δt) before the cosine time encoding to balance skewed time differences, and explicitly maintain a per-node-pair history of recent edge embeddings fed to the link predictor, so that TGNNs can be plugged in with minimal changes.
Key results
- TAMI improves GraphMixer and DyGFormer AP on all 13 classic datasets under random negative sampling, with gains up to 16.64% for GraphMixer.
- On TGB tgbl-review, DyGFormer with TAMI improves MRR from 0.224 to 0.419 (87.05%), ranking second.
- TAMI lets DyGFormer on UC and EN use 2 and 32 historical interactions instead of 16x and 8x more, cutting GPU memory by up to 53.22%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TAMI (LTE and LHA) to DyGFormer | DyGFormer · tgbl-review | MRR: 0.224 -> 0.419; 87.05% | Helped |
| Add TAMI (LTE and LHA) to DyGFormer | DyGFormer · rnd temporal link prediction, 13 classic datasets | AP | Helped |
| Add LTE alone (ablation) | GraphMixer · LA | AP: 75.56 -> 75.21 | Hurt |
| Add LTE alone (ablation) | GraphMixer · UC | AP: 93.38 -> 94.98 | Helped |
| Add LTE alone (ablation) | DyGFormer · UT | AP: 65.07 -> 67.24 | Helped |
| Add LHA alone (ablation) | GraphMixer · EN | AP: 82.26 -> 89.88 | Helped |
| Add LHA alone (ablation) | DyGFormer · CP | AP: 97.91 -> 97.57 | Hurt |
| Add LHA alone (ablation) | DyGFormer · UT | AP: 65.07 -> 64.53 | Hurt |
| Add LTE and LHA together (ablation) | GraphMixer · UC | AP: 93.38 -> 96.20 | Helped |
| Add LTE and LHA together (ablation) | DyGFormer · UC | AP: 95.66 -> 96.72 | Helped |
| Use LTE's logarithmic time-difference transformation (ablation) | GraphMixer on UCI · UCI | temporal-difference skewness: 2.385 to -1.14 | Helped |
| Use 2 historical interactions for attention instead of 16× more | DyGFormer on UC · UC | GPU memory usage: up to 53.22% | Helped |
Surprising findings
- LTE alone slightly hurts GraphMixer on a few datasets (e.g., LA -0.46%, SE -0.75%, US -0.94%).
- LHA alone slightly reduces DyGFormer performance on some datasets (e.g., UT -0.83%, CP -0.35%).
Limitations
- Gains are reported mainly for two base TGNNs (GraphMixer and DyGFormer) and some models show inconsistent per-component gains.
- Proposition 1 holds only under a Pareto assumption; real data is not strictly Pareto.
Experimental setup
Datasets: Can. Parl. Contact Enron Flights Lastfm Mooc Reddit Social Evo Uci UN Trade UN Vote US Legis
Benchmarks: Can. Parl. Contact Enron Flights Lastfm Mooc Reddit Social Evo Uci UN Trade UN Vote US Legis
Metrics: average precision mean reciprocal rank
Baselines: JODIE DyRep TGAT TGN CAWN EdgeBank TCL GraphMixer DyGFormer
Builds on: GraphMixer DyGFormer TGAT TGN CAWN
Closest papers by meaning
- Improving Temporal Link Prediction via Temporal Walk Matrix Projection (NeurIPS 2024 · similarity 0.90)
- Future Link Prediction Without Memory or Aggregation (NeurIPS 2025 · similarity 0.90)
- S3Former: Sequential, Structural, and Statistical Fusion for Continuous-Time Dynamic Graphs (NeurIPS 2026 · similarity 0.88)
- TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · similarity 0.87)
- Using Time-Aware Graph Neural Networks to Predict Temporal Centralities in Dynamic Graphs (NeurIPS 2024 · similarity 0.87)
- TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics (ICLR 2025 · similarity 0.86)
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.