Temporal link prediction uses past interactions to predict future links. In practice, interaction records from web and mobile applications can contain gaps caused by logging failures, API limits or lost records, even when later events remain available. To our knowledge, the extent to which later events help recover interactions within such gaps has not been measured directly for this ranking task. We study retrospective reconstruction using the tgbl-wiki and tgbl-comment datasets. We hide a two-hour window, use its interactions as ground truth, and rank the directed pairs active within it. We compare count-based heuristics and logistic regression models using past information alone or together with 1, 6 or 24 hours of later events. Target pairs and candidate lists remain fixed across conditions, so differences in ranking performance cannot be attributed to changes in candidate selection. With 24 hours of later events, the 26-feature logistic regression improves mean reciprocal rank (MRR) by 0.039 on tgbl-wiki under exhaustive ranking, 0.046 under historical-random ranking, and 0.108 on tgbl-comment under historical-random ranking. The 26-feature logistic regression benefits from adding later features to past features, whereas replacing the past pair count with the later pair count reduces MRR by 0.157 on tgbl-wiki under exhaustive ranking. A control removing later interactions from query sources to historically observed destinations reduces the MRR gain by 59% on tgbl-wiki and 10% on tgbl-comment. Because this control removes more than repeated target pairs, these reductions cannot be attributed to pair recurrence alone. Later events improve the ranking of hidden interactions on both datasets, including tgbl-comment, where only 6.4% of target pairs reappeared within 24 hours.
Figures & tables
Figure 1: Episode construction. The hidden window is excluded from all model inputs, and target pairs are identified from the events within that window.
Protocol
Construction
Role
Exhaustive
Every destination in the historical universe
Primary for tgbl-wiki , not feasible for tgbl-comment
Historical-random
100 negatives: up to 50 past destinations of the query source, with the remainder sampled uniformly
Primary for tgbl-comment , matched comparison across datasets
Degree-matched
100 negatives from the same past in-degree bin as the true destination, with four bins fitted on training episodes
Sensitivity analysis
Table 1: Candidate protocols.
Figure 2: Information conditions and controls. All conditions share the same past history and hidden-window targets.
Dataset
Protocol
Past MRR
Past + 24 h
Gain 1 h
Gain 6 h
Gain, 24 h [95% interval]
tgbl-wiki
exhaustive
0.599
0.639
+0.015
+0.029
+0.039 [0.022, 0.058]
tgbl-wiki
hist.-random
0.617
0.663
+0.018
+0.040
+0.046 [0.029, 0.064]
tgbl-comment
hist.-random
0.485
0.593
+0.050
+0.096
+0.108 [0.098, 0.118]
Table 2: Test MRR of the 26-feature logistic regression with past-only and past-plus-later information. Gains are seed-averaged, and the 24-hour gain is shown with its 95% paired user-cluster bootstrap interval. Hist.-random denotes historical-random ranking. Gains are computed from unrounded values, so they can differ by 0.001 from the difference of the displayed MRRs.
Figure 3: Gain in test MRR from adding later events, for the 26-feature logistic regression at each horizon. Points are seed-averaged differences between the past-plus-later and past-only conditions, and bars are 95% paired user-cluster bootstrap intervals.
Model family
tgbl-wiki exhaustive
tgbl-wiki hist.-random
tgbl-comment hist.-random
Pair heuristic (later count replaces past count)
−0.157
−0.140
+0.064
Popularity heuristic
+0.006
+0.005
+0.009
Logistic regression, 6 activity features
+0.007
+0.056
+0.121
Logistic regression, 26 features
+0.039
+0.046
+0.108
Table 3: Change in test MRR at 24 hours from adding later events, by model family.
Figure 4: Full gain from later events and the MRR cost of the two controls at 24 hours under historical-random ranking. Bars are 95% paired user-cluster bootstrap intervals. The shares are ratios of point estimates.
Figure 5: Gain in test MRR at 24 hours by group, under the primary protocol of each dataset. Groups are defined by later outcomes and are descriptive. The tgbl-wiki group of new pairs that appeared again holds only 9 lists, and its estimate is highly uncertain.
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
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.
Temporal link prediction on the Temporal Graph Benchmark 2.0 (TGB 2.0) faces a scalability ceiling: on the benchmark's three largest datasets, every existing embedding method runs out of memory or exceeds the time budget. These large-scale graphs are the ones nearest real deployment scale, so failing on them is a real production limitation. EdgeReMIND sets the highest reported test mean reciprocal rank (MRR) on six of eight TGB 2.0 datasets and is the only relation-aware method that runs on all of them. This linear memorization model, with learned per-relation weights over data-calibrated features, is therefore not merely a fallback where embeddings fail but a practical state-of-the-art baseline across the benchmark.