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
Fig. 1: Overview of StreamTab. Interaction histories provide tabular features for context examples and query pairs. Context negatives follow the sampling regime, and a frozen TFM scores query pairs using recent labeled examples.
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.
TABLE I: Historical statistics used to represent a candidate interaction. Source and destination features are computed separately for each endpoint role.
Dataset
Domain
Nodes
Interactions
Unique pairs
Bip.
Surprise
Wikipedia
Social
9,227
157,474
18,257
Yes
0.291
Reddit
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
TABLE II: Statistics of the temporal graph datasets. Interactions include repeated occurrences of a pair; unique pairs count distinct source–destination pairs. Bip. indicates a bipartite graph; Surprise is the fraction of test interactions involving pairs absent from the training and validation periods.
Method
Wiki.
Reddit
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
TABLE III: Full-test AP (%) under historical negative sampling. Bold and underline mark the best and second-best means at the displayed precision; ties share the same mark. Wins / 13 counts first places. A dash indicates an unavailable result.
Method
Wiki.
Reddit
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
TABLE IV: Full-test AP (%) under inductive negative sampling. Conventions follow Table III .
Fig. 2: Effect of regime conditioning. Left: AP differences between regime-matched and random-negative contexts, in percentage points (pp); large markers show means over three runs, small markers show individual runs, and bars show their range. Right: mean AP (%) for each combination of test sampling (rows) and context sampling (columns). Outlined cells mark the best context for each test setting.
Sampling regime
Wins / 13
Median Δ AP (pp)
Random
12
+0.64
Historical
10
+0.13
Inductive
13
+2.28
TABLE V: Predictor comparison with the same inputs. Top: TabPFN versus LightGBM across thirteen datasets. Wins count datasets where TabPFN achieves higher mean AP than LightGBM; Δ AP denotes TabPFN minus LightGBM, measured in percentage points (pp). Below: AP (%) on four datasets (mean ± standard deviation over three runs). Bold marks the best mean.
Fig. 3: AP change and evaluation speedup across fifteen context configurations on MOOC and LastFM. Each line represents a context size k , with points for R=1,4,8,16,32 . AP differences (percentage points) and evaluation speedups are relative to k=8,192 , R=1 for the same dataset and sampling regime, using means over three runs.
Dataset
Regime
Fixed
Refresh
Δ 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
TABLE VI: AP (%) with periodic refresh and a fixed initial context (mean ± standard deviation over three matched runs). Refresh uses R=4 ; both variants use k=8,192 . Δ AP is the mean paired difference (refresh minus fixed), in percentage points.
Fig. 4: AP over the test period under random sampling. Curves show mean AP over three runs with matched seeds, smoothed with a centered 25-batch moving average and shorter windows at the boundaries.
Fig. 5: AP difference between StreamTab and the best published baseline versus test surprise under three negative sampling regimes. Positive values favor StreamTab; each point represents one dataset.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
x
p+(x)
p−(r)(x)
p−(s)(x)
ηr(x)
ηs(x)
a
1/2
4/5
1/5
5/13
5/7
b
1/2
1/5
4/5
5/7
5/13
Appendix
TABLE S1: A constructed example in which two negative distributions reverse the ordering of two candidates.
Method
Wiki.
Reddit
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
Appendix
TABLE S2: Full-test AP (%) under random negative sampling. Values are means; bold and underline mark the best and second-best displayed means, and a dash indicates an unavailable published result. Wins / 13 counts first places, including ties at the displayed precision.
Historical
Inductive
Random
Dataset
Best baseline
StreamTab
Δ
Best baseline
StreamTab
Δ
Best baseline
StreamTab
Δ
Wikipedia
GraphMixer 87.60
90.59
+2.99
CTAN 92.61
75.98
− 16.63
FreeDyG 98.97
97.70
− 1.27
Reddit
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
Appendix
TABLE S3: New-node AP (%): the best published result and StreamTab under each regime. Values are means; Δ is StreamTab minus the published result in percentage points. Bold marks a StreamTab lead; Wins / 13 counts its first places, including ties at the displayed precision.
Method
Wikipedia
Reddit
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
Appendix
TABLE S4: Full-test AP (%) under historical negative sampling, reported as mean ± standard deviation. Bold and underline mark the best and second-best displayed means; a dash indicates an unavailable result. StreamTab uses three runs; baseline statistics are as published.
Method
Wikipedia
Reddit
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
Appendix
TABLE S5: Full-test AP (%) under inductive negative sampling, reported as mean ± standard deviation. Conventions follow Table S4 .
Method
Wikipedia
Reddit
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
Appendix
TABLE S6: Full-test AP (%) under random negative sampling, reported as mean ± standard deviation. Conventions follow Table S4 .
Method
Wikipedia
Reddit
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
Appendix
TABLE S7: Full-test AUC (%) under historical negative sampling, reported as mean ± standard deviation. Conventions follow Table S4 .
Method
Wikipedia
Reddit
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
Appendix
TABLE S8: Full-test AUC (%) under inductive negative sampling, reported as mean ± standard deviation. Conventions follow Table S4 .
Method
Wikipedia
Reddit
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
Appendix
TABLE S9: Full-test AUC (%) under random negative sampling, reported as mean ± standard deviation. Conventions follow Table S4 .
Random
Wiki.
Reddit
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
Appendix
TABLE S10: AP (%) of predictors given the same features and matched context. Values are means over three runs; bold marks the best displayed mean for each dataset and regime. TabPFN and LightGBM cover all thirteen datasets; HGB and TabICL cover Wikipedia, Reddit, MOOC, LastFM, Enron, and UCI. A dash indicates a predictor that was not evaluated on that dataset.
Fig. S1: AP over the test period under historical (left) and inductive (right) negative sampling. Curves show mean AP over three runs with matched seeds, smoothed with a centered 25-batch moving average and shorter windows at the boundaries.
Fig. S2: Change in AP (percentage points) against evaluation speedup on six datasets under random, historical, and inductive negative sampling. Each line represents a context size k , with points for R=1,4,8,16,32 . AP differences and evaluation speedups are relative to k=8,192 , R=1 for the same dataset and sampling regime, using means over three runs.
Many temporal link predictors summarize past interactions through learned node representations. We examine whether simple counts of recurring interaction patterns can provide competitive predictions without learning these representations. We propose a temporal link predictor based on statistical language modelling. It pools transition and co-occurrence counts across sources to predict links that a source has never formed. We smooth sparse estimates using destination frequencies or Kneser-Ney continuation counts. A shared log-linear rule combines these estimates with popularity, source history, and recency, without node embeddings. In our main evaluation, the model achieves the highest MRR among the compared methods on 7 out of 16 datasets from TGB and TGB-Seq. It also outperforms EdgeBank and Base3 on all 16 datasets and the heuristic family on 14. These gains extend to datasets designed to limit repeated edges. With only 9--13 learned parameters, our model provides a simple and competitive baseline for evaluating future neural temporal link predictors.
Lisi Qarkaxhija, Ingo Scholtes
Chair of Machine Learning for Complex Networks Center for Artificial Intelligence and Data Science (CAIDAS) Julius-Maximilians-Universität Würzburg, DE
Existing temporal link prediction methods typically rely on learned temporal representations to capture evolving interaction patterns. In this paper, we show that a classical structural proximity measure can match or outperform these methods without training or learned temporal sequence modeling. We revisit Personalized PageRank (PPR) on historical interaction graphs, using simple sliding-window retention or exponential time decay. Surprisingly, across 14 datasets, Time-decay PPR ranks first on nine and second on four among the compared methods, including datasets with few repeated edges. To understand this result, we examine the historical paths connecting source nodes to previously unseen destinations. We find that such paths contain useful predictive information, and that PPR outperforms a strong neural baseline on queries where historical structure clearly distinguishes the true destination from negative candidates. These results show that strong performance on current benchmarks does not necessarily require complex temporal models. More fundamentally, they raise the question of how much predictive value learned temporal modeling adds beyond historical structural proximity. Our findings motivate reassessing progress in temporal link prediction with strong classical baselines.
An interaction has two parts: someone decides to act, and then chooses whom to act on. Temporal link prediction has concentrated on the second, and we show that it is blind to the first by construction: a standard negative keeps the real source and swaps the destination, and we prove that this cancels the source's activity exactly from the optimal score, so no model trained and evaluated this way is ever rewarded for learning it. Under the harder historical and inductive negatives, whose sources differ, the same factor becomes the dominant signal. We model it with Source Node Activity Modeling (SNAM), a self-exciting event intensity fitted by an exact point-process likelihood to decayed interaction counts the history states already contain; it has fewer than 20 parameters. On their own, never looking at the destination, these parameters beat DyGFormer and TPNet on four of five datasets under historical negatives. Added to the frozen scores of TPNet, TGN, DyGFormer and DSRD, four models of different design, without retraining anything, they raise AP on almost every backbone-dataset pair in both settings, by up to 25 points. Our full model ranks first overall against eleven baselines on 13 datasets and three protocols, and on million-event streams trains an epoch 9-100x faster than TPNet and DyGFormer. We conclude that source activity is a blind spot of temporal link prediction, and a cheap, transferable one to close. Code is available at https://github.com/Erutaner/Your-Temporal-Link-Predictor-Is-Blind-to-Who-Is-Active.