Smoothed-Count Baseline for Temporal Link Prediction without Learned Memory

What happened

The research team challenges the dominant temporal link prediction paradigm that relies on learned node representations, and proposes a new predictor built on statistical language modeling: it aggregates transition and co-occurrence counts, applies sparse estimation smoothing with destination frequency or Kneser-Ney continuation counts, fuses popularity, source history and recency signals via a shared log-linear rule, and uses no node embeddings throughout. Evaluated across 16 datasets from TGB and TGB-Seq, the model achieves the highest MRR on 7 datasets, outperforms EdgeBank and Base3 across the full dataset suite, beats heuristic methods on 14 datasets, and requires only 9-13 learnable parameters.

Key facts

Number of learnable parameters

9-13

Evaluation benchmarks

16 datasets across TGB and TGB-Seq

Core strengths

Top MRR on 7 datasets, outperforms EdgeBank and Base3 across the full dataset suite

Background

Most prior temporal link prediction work relies on learned node representations to summarize historical interactions, with generally high model complexity.

Why it matters

This finding breaks the field's path dependence on complex neural representation learning for temporal link prediction: for researchers, the strong low-parameter baseline drastically lowers the reference bar for evaluating new models, avoiding misleading results from redundant parameters in overcomplicated models; for engineering practitioners, it delivers competitive predictive performance without training large models, significantly reducing deployment costs.

What to watch

This baseline will be used to evaluate future neural temporal link prediction methods as a new reference benchmark.