Temporal Motif Signatures for Temporal Graph Neural Networks

Dylan Sandfelder, Mihai Cucuringu, Xiaowen Dong · 2026

Read Temporal Motif Signatures for Temporal Graph Neural Networks by Dylan Sandfelder, Mihai Cucuringu, Xiaowen Dong (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-01176.txt.

Description

Sandfelder, Cucuringu, and Dong hand graph networks a small set of motif counts in this 2026 paper. Temporal interaction streams hide signal in short patterns: repeats, replies, star shapes, and three way flows. Plain temporal graph networks often fail to show those patterns to their edge scorers. The structure sits in the data, unused.

On MOOC interaction prediction, a four feature family of past window star counts already takes most of the lift over a strong static baseline. Across many real and synthetic sets, motif activity sorts along three stable axes. They cover dyadic recency and reciprocity, star diversity, and triadic flow. The team builds a compact map of 13 motif coordinates per candidate edge. It is leakage safe and local, and it plugs into any encoder with no architecture change. Tests span link prediction, edge classification, and graph classification. The same features lift scores across all of them, they report.

Catalog record

CREATORDylan Sandfelder, Mihai Cucuringu, Xiaowen Dong
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.01176

Browse the full library · About the Tower of Babel