Dynamic-graph prediction without per-graph retraining: interaction histories become tables a frozen tabular model can rank
Synopsis
CL-DGFM encodes interaction histories on dynamic graphs as tabular features independent of node identity, covering repeated contacts, indirect connectivity and ordered destination transitions, then lets a pretrained tabular foundation predictor learn from diverse synthetic dynamics how to read them. It achieves the highest mean MRR on 10 of 13 datasets across TGB and TGB-Seq, exceeding the strongest listed baselines by 10.78, 15.24 and 11.69 percentage points on Comment, Flickr and YouTube, and transfers to real graphs without target-graph parameter updates.
Interpretation
Temporal link prediction on dynamic graphs can now reuse one predictor instead of training a separate model for each network. Existing temporal graph models are typically trained separately per dataset, so a network with different entities and activity patterns requires new training. Across 13 datasets from TGB and TGB-Seq, CL-DGFM achieves the highest mean MRR on 10, exceeding the strongest listed baselines by 10.78, 15.24 and 11.69 percentage points on Comment, Flickr and YouTube respectively.
The route is to write each candidate interaction as one row of numeric features and hand it to an off-the-shelf tabular predictor for scoring. A dynamic interaction tokenizer uses only history before the query time, compressing direct-contact memory, indirect connectivity through historical paths and ordered destination transitions into features calibrated by within-list ranks, without encoding node identities. Once stacked into tables, TabICLv2, TabPFNv2 and LimiX-2M all predict with fixed parameters and only the target graph's labeled historical context, requiring no additional trainable graph encoder.
On the training side, synthetic interaction streams cover many relationships between history and future outcomes so the predictor learns to read the same features under different dynamics. Learning from a limited range of interaction processes can favor one historical signal and restrict transfer to networks where other signals matter. At a fixed update budget, expanding the synthetic graph pool from 256 to 16,384 raises MRR on all five real graphs, from 68.56 to 82.18 on Wiki and from 62.22 to 76.02 on Flickr.
Perspective
The demonstrated scope is temporal link prediction on 13 datasets from TGB and TGB-Seq, with the model trained only on synthetic interaction streams, parameters fixed at inference, and 512 positive events sampled from the final 10% of each graph's training interval serving as labeled context. It is aimed at researchers working on graph foundation models and tabular foundation model reuse.
CL-DGFM remains slightly below TPNet on Wiki and shows larger gaps on GoogleLocal and Patent; more context is not uniformly better, as frozen TabICLv2 on Wiki declines as context grows toward 512 positive events before partially recovering.