StreamTab: Training-Free Temporal Link Prediction with Tabular Foundation Models
Organizations: South China University of Technology, Guangzhou, Guangdong, China
Abstract
Temporal link prediction uses past interactions to predict whether two nodes will interact in the future. Temporal graph neural networks achieve strong results on this task, but they are trained separately for each graph. How a model should use interaction history also depends on the candidates it is compared with. A past interaction can distinguish a recurring link from a random pair that has never interacted, but when the negative candidates have also interacted before, other signals are needed. We propose StreamTab, a training-free framework built on a frozen tabular foundation model pretrained only on synthetic tables. StreamTab describes each candidate pair by statistics of its past interactions and shows the model recent observed interactions and sampled negative pairs as labeled examples, from which the model scores new candidates. Its central design, regime conditioning, samples the negative examples according to how the candidates are sampled, so the examples show the distinction the model must make. As new interactions arrive, StreamTab updates the statistics and refreshes the examples while the model parameters stay fixed. On thirteen datasets against 21 temporal graph baselines, StreamTab achieves the highest average precision on twelve datasets under historical sampling and eleven under inductive sampling, and ranks first or stays within one point of the best on seven datasets under random sampling. With the same features and model, regime conditioning raises average precision over examples with random negatives by a median of 12.6 points under historical sampling.
Figures & tables
| Group | Features | Historical information |
|---|---|---|
| Pair | 7 | Recency, total and windowed counts, time since first interaction, mean recurrence gap, and elapsed time relative to that gap. |
| Source | 9 | Recency, total and windowed activity, distinct partners, recent partner novelty, and interaction-gap statistics. |
| Destination | 9 | The corresponding statistics in the destination role. |
| Popularity | 4 | Destination activity shares at two scales, their log ratio, and the source activity share. |
| Neighborhood | 4 | Endpoint appearances in recent partner histories, common-neighbor count, and neighborhood Jaccard similarity. |
| Global | 1 | Time since the last event in the history. |
| Dataset | Domain | Nodes | Interactions | Unique pairs | Bip. | Surprise |
|---|---|---|---|---|---|---|
| Wikipedia | Social | 9,227 | 157,474 | 18,257 | Yes | 0.291 |
| Social | 10,984 | 672,447 | 78,516 | Yes | 0.111 | |
| MOOC | Interaction | 7,144 | 411,749 | 178,443 | Yes | 0.796 |
| LastFM | Interaction | 1,980 | 1,293,103 | 154,993 | Yes | 0.419 |
| Enron | Social | 184 | 125,235 | 3,125 | No | 0.269 |
| Social Evo. | Proximity | 74 | 2,099,519 | 4,486 | No | 0.002 |
| Method | Wiki. | MOOC | LastFM | Enron | S.Evo | UCI | Flights | CanP. | USLeg. | UNTr. | UNVo. | Cont. | Wins / 13 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JODIE | 83.01 | 80.03 | 78.94 | 74.35 | 69.85 | 87.44 | 75.24 | 66.48 | 51.79 | 51.71 | 61.39 | 70.02 | 95.31 | 0 |
| DyRep | 79.93 | 79.83 | 75.60 | 74.92 | 71.19 | 93.29 | 55.10 | 67.61 | 63.31 | 86.88 | 59.19 | 69.30 | 96.39 | 0 |
| TGN | 86.86 | 81.22 | 87.06 | 76.87 | 73.91 | 94.45 | 80.43 | 66.70 | 68.42 | 74.00 | 58.44 | 69.37 | 93.05 | 0 |
| CTAN | 95.91 | 89.77 | 67.73 | 82.29 | 77.24 | — | 76.62 | 75.75 | 89.28 | 84.20 | 67.38 | 74.75 | 97.86 | 0 |
| CAWN | 71.21 | 80.82 | 74.05 | 69.86 | 64.73 | 85.53 | 65.30 | 64.72 | 66.53 | 68.82 | 55.71 | 51.26 | 84.16 | 0 |
| NAT | 68.36 | 80.58 | 85.42 | 78.75 | 72.90 | 91.29 | 75.38 | 64.74 | 77.72 | 91.12 | 78.65 | 71.39 | 96.84 | 0 |
| Method | Wiki. | MOOC | LastFM | Enron | S.Evo | UCI | Flights | CanP. | USLeg. | UNTr. | UNVo. | Cont. | Wins / 13 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JODIE | 75.65 | 86.98 | 65.23 | 62.67 | 68.96 | 89.82 | 65.99 | 69.07 | 48.42 | 50.27 | 60.42 | 67.79 | 93.43 | 0 |
| DyRep | 70.21 | 86.30 | 61.66 | 64.41 | 67.79 | 93.28 | 54.79 | 70.57 | 58.61 | 83.44 | 60.19 | 67.53 | 94.18 | 0 |
| TGN | 85.62 | 88.10 | 77.50 | 65.95 | 70.89 | 95.13 | 70.94 | 71.09 | 65.34 | 67.57 | 61.04 | 67.63 | 90.18 | 0 |
| CTAN | 94.15 | 90.99 | 64.93 | 80.06 | 72.02 | — | 66.25 | 76.20 | 87.42 | 79.13 | 68.98 | 72.85 | 97.31 | 0 |
| CAWN | 74.06 | 91.67 | 73.51 | 67.48 | 75.15 | 88.32 | 64.61 | 69.18 | 67.75 | 65.81 | 62.54 | 52.19 | 89.31 | 0 |
| NAT | 53.85 | 81.71 | 77.28 | 64.02 | 70.34 | 92.12 | 57.55 | 59.16 | 78.03 | 88.40 | 78.38 | 72.53 | 91.13 | 0 |
| Sampling regime | Wins / 13 | Median AP (pp) |
|---|---|---|
| Random | 12 | |
| Historical | 10 | |
| Inductive | 13 |
| Dataset | Regime | Fixed | Refresh | AP |
|---|---|---|---|---|
| UCI | Random | 89.12 1.68 | 91.44 0.27 | +2.31 |
| Historical | 98.18 0.01 | 98.34 0.01 | +0.16 | |
| Inductive | 90.02 0.30 | 90.73 0.09 | +0.71 | |
| MOOC | Random | 81.48 0.15 | 83.68 0.08 | +2.20 |
| Historical | 98.45 0.00 | 98.48 0.01 | +0.03 | |
| Inductive | 97.43 0.07 | 98.21 0.05 | +0.78 |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Wiki. | MOOC | LastFM | Enron | S.Evo | UCI | Flights | CanP. | USLeg. | UNTr. | UNVo. | Cont. | Wins / 13 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JODIE | 96.50 | 98.31 | 80.23 | 70.85 | 84.77 | 89.89 | 89.43 | 95.60 | 69.26 | 75.05 | 64.94 | 63.91 | 95.31 | 0 |
| DyRep | 94.86 | 98.22 | 81.97 | 71.92 | 82.38 | 88.87 | 65.14 | 95.29 | 66.54 | 75.34 | 63.21 | 62.81 | 95.98 | 0 |
| TGN | 98.45 | 98.63 | 89.15 | 77.07 | 86.53 | 93.57 | 92.34 | 97.95 | 70.88 | 75.99 | 65.03 | 65.72 | 96.89 | 0 |
| CTAN | 96.61 | 97.21 | 84.71 | 86.44 | 92.52 | — | 76.64 | 93.64 | 87.01 | 79.44 | 69.01 | 70.19 | 98.53 | 0 |
| CAWN | 98.76 | 99.11 | 80.15 | 86.99 | 89.56 | 84.96 | 95.18 | 98.51 | 69.82 | 70.58 | 65.39 | 52.84 | 90.26 | 0 |
| NAT | 98.03 | 99.13 | 85.88 | 88.02 | 90.60 | 88.92 | 93.40 | 98.57 | 79.72 | 78.71 | 73.95 | 70.45 | 97.39 | 0 |
| Historical | Inductive | Random | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Dataset | Best baseline | StreamTab | Best baseline | StreamTab | Best baseline | StreamTab | |||
| Wikipedia | GraphMixer 87.60 | 90.59 | +2.99 | CTAN 92.61 | 75.98 | 16.63 | FreeDyG 98.97 | 97.70 | 1.27 |
| TAWRMAC 67.95 | 83.57 | +15.62 | CTAN 78.35 | 65.38 | 12.97 | TAWRMAC 98.99 | 98.74 | 0.25 | |
| MOOC | TIDFormer 84.61 | 95.21 | +10.60 | TIDFormer 84.61 | 90.91 | +6.30 | TPNet 95.07 | 83.56 | 11.51 |
| LastFM | TIDFormer 82.79 | 88.94 | +6.15 | TIDFormer 82.79 | 54.83 | 27.96 | TIDFormer 95.41 | 93.61 | 1.80 |
| Enron | RepeatMixer 84.49 | 74.17 | 10.32 | RepeatMixer 84.48 | 68.47 | 16.01 | TPNet 90.34 | 89.85 | 0.49 |
| Method | Wikipedia | MOOC | LastFM | Enron | Social Evo. | UCI | |
|---|---|---|---|---|---|---|---|
| JODIE | 83.01 0.66 | 80.03 0.36 | 78.94 1.25 | 74.35 3.81 | 69.85 2.70 | 87.44 6.78 | 75.24 5.80 |
| DyRep | 79.93 0.56 | 79.83 0.31 | 75.60 1.12 | 74.92 2.46 | 71.19 2.76 | 93.29 0.43 | 55.10 3.14 |
| TGN | 86.86 0.33 | 81.22 0.61 | 87.06 1.93 | 76.87 4.64 | 73.91 1.76 | 94.45 0.56 | 80.43 2.12 |
| CTAN | 95.91 0.10 | 89.77 2.28 | 67.73 2.08 | 82.29 0.94 | 77.24 1.53 | — | 76.62 0.33 |
| CAWN | 71.21 1.67 | 80.82 0.45 | 74.05 0.95 | 69.86 0.43 | 64.73 0.36 | 85.53 0.38 | 65.30 0.43 |
| NAT | 68.36 4.47 | 80.58 0.91 | 85.42 3.26 | 78.75 1.63 | 72.90 1.38 | 91.29 4.89 | 75.38 1.28 |
| Method | Wikipedia | MOOC | LastFM | Enron | Social Evo. | UCI | |
|---|---|---|---|---|---|---|---|
| JODIE | 75.65 0.79 | 86.98 0.16 | 65.23 2.19 | 62.67 4.49 | 68.96 0.98 | 89.82 4.11 | 65.99 1.40 |
| DyRep | 70.21 1.58 | 86.30 0.26 | 61.66 0.95 | 64.41 2.70 | 67.79 1.53 | 93.28 0.48 | 54.79 1.76 |
| TGN | 85.62 0.44 | 88.10 0.24 | 77.50 2.91 | 65.95 5.98 | 70.89 2.72 | 95.13 0.56 | 70.94 0.71 |
| CTAN | 94.15 0.08 | 90.99 2.19 | 64.93 3.31 | 80.06 0.85 | 72.02 2.64 | — | 66.25 0.51 |
| CAWN | 74.06 2.62 | 91.67 0.24 | 73.51 0.94 | 67.48 0.77 | 75.15 0.58 | 88.32 0.27 | 64.61 0.48 |
| NAT | 53.85 6.25 | 81.71 1.03 | 77.28 1.94 | 64.02 1.48 | 70.34 1.73 | 92.12 4.53 | 57.55 0.81 |
| Method | Wikipedia | MOOC | LastFM | Enron | Social Evo. | UCI | |
|---|---|---|---|---|---|---|---|
| JODIE | 96.50 0.14 | 98.31 0.14 | 80.23 2.44 | 70.85 2.13 | 84.77 0.30 | 89.89 0.55 | 89.43 1.09 |
| DyRep | 94.86 0.06 | 98.22 0.04 | 81.97 0.49 | 71.92 2.21 | 82.38 3.36 | 88.87 0.30 | 65.14 2.30 |
| TGN | 98.45 0.06 | 98.63 0.06 | 89.15 1.60 | 77.07 3.97 | 86.53 1.11 | 93.57 0.17 | 92.34 1.04 |
| CTAN | 96.61 0.79 | 97.21 0.84 | 84.71 2.85 | 86.44 0.80 | 92.52 1.20 | — | 76.64 4.11 |
| CAWN | 98.76 0.03 | 99.11 0.01 | 80.15 0.25 | 86.99 0.06 | 89.56 0.09 | 84.96 0.09 | 95.18 0.06 |
| NAT | 98.03 0.07 | 99.13 0.10 | 85.88 0.55 | 88.02 1.94 | 90.60 0.66 | 88.92 3.45 | 93.40 0.26 |
| Method | Wikipedia | MOOC | LastFM | Enron | Social Evo. | UCI | |
|---|---|---|---|---|---|---|---|
| JODIE | 80.77 0.73 | 80.52 0.32 | 82.75 0.83 | 75.22 2.36 | 75.39 2.37 | 90.06 3.15 | 78.64 3.50 |
| DyRep | 77.74 0.33 | 80.15 0.18 | 81.06 0.94 | 74.65 1.98 | 74.69 3.55 | 93.12 0.34 | 57.91 3.12 |
| TGN | 82.74 0.32 | 81.11 0.19 | 88.00 1.80 | 77.97 3.04 | 77.09 2.22 | 94.71 0.53 | 77.25 2.68 |
| CTAN | 95.43 0.07 | 90.63 2.28 | 73.87 2.77 | 79.50 0.82 | 81.95 1.64 | — | 75.05 0.13 |
| CAWN | 67.84 0.64 | 80.27 0.30 | 71.57 1.07 | 67.88 0.24 | 65.10 0.34 | 87.43 0.15 | 57.86 0.15 |
| NAT | 69.32 2.78 | 79.36 0.31 | 84.93 2.89 | 75.89 2.21 | 73.22 2.18 | 91.87 4.52 | 71.52 1.61 |
| Method | Wikipedia | MOOC | LastFM | Enron | Social Evo. | UCI | |
|---|---|---|---|---|---|---|---|
| JODIE | 70.96 0.78 | 83.51 0.15 | 66.63 2.30 | 61.32 3.49 | 70.92 1.05 | 90.01 3.19 | 64.14 1.26 |
| DyRep | 67.36 0.96 | 82.90 0.31 | 63.26 1.01 | 62.15 2.12 | 68.73 1.34 | 93.07 0.38 | 54.25 2.01 |
| TGN | 80.97 0.31 | 84.56 0.24 | 77.44 2.86 | 65.46 4.27 | 71.34 2.46 | 95.24 0.56 | 64.11 1.04 |
| CTAN | 93.67 0.11 | 91.42 2.18 | 68.17 3.73 | 78.70 0.87 | 75.40 1.92 | — | 66.51 0.25 |
| CAWN | 70.95 0.95 | 88.04 0.29 | 70.32 1.43 | 67.92 0.44 | 75.17 0.50 | 89.93 0.15 | 58.06 0.26 |
| NAT | 52.11 5.83 | 76.01 1.04 | 75.56 2.39 | 58.54 2.37 | 69.62 1.12 | 92.53 4.20 | 49.73 1.20 |
| Method | Wikipedia | MOOC | LastFM | Enron | Social Evo. | UCI | |
|---|---|---|---|---|---|---|---|
| JODIE | 96.33 0.07 | 98.31 0.05 | 83.81 2.09 | 70.49 1.66 | 87.96 0.52 | 92.05 0.46 | 90.44 0.49 |
| DyRep | 94.37 0.09 | 98.17 0.05 | 85.03 0.58 | 71.16 1.89 | 84.89 3.00 | 90.76 0.21 | 68.77 2.34 |
| TGN | 98.37 0.07 | 98.60 0.06 | 91.21 1.15 | 78.47 2.94 | 88.32 0.99 | 95.39 0.17 | 92.03 1.13 |
| CTAN | 97.00 0.21 | 97.24 0.75 | 85.40 2.67 | 85.12 0.77 | 87.09 1.51 | — | 76.25 2.83 |
| CAWN | 98.54 0.04 | 99.01 0.01 | 80.38 0.26 | 85.92 0.10 | 90.45 0.14 | 87.34 0.08 | 93.87 0.08 |
| NAT | 97.75 0.11 | 99.09 0.10 | 87.42 0.58 | 86.92 2.72 | 91.68 0.83 | 90.84 3.72 | 92.31 0.37 |
| Random | Wiki. | MOOC | LastFM | Enron | S.Evo | UCI | Flights | CanP. | USLeg. | UNTr. | UNVo. | Cont. | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LightGBM | 98.12 | 98.95 | 82.44 | 91.75 | 93.19 | 93.31 | 82.49 | 96.63 | 85.56 | 78.12 | 85.60 | 76.93 | 98.58 |
| HGB | 97.91 | 98.95 | 81.75 | 91.61 | 92.84 | — | 83.81 | — | — | — | — | — | — |
| TabICL | 98.62 | 99.15 | 82.43 | 92.36 | 93.51 | — | 83.49 | — | — | — | — | — | — |
| TabPFN | 98.63 | 99.16 | 83.68 | 92.37 | 93.83 | 94.11 | 91.44 | 98.55 | 84.33 | 78.94 | 85.96 | 77.74 | 98.79 |
Temporal Link Prediction
Temporal link prediction forecasts which relationships will form or recur as networks and knowledge graphs change over time.
See all 67 papers