Retour au feed
arXiv cs.LG·

Spatiotemporal Imputation with Graph-Informed Flow Matching

Signal
75
Hype
15
En 3 lignesGiFlow, un framework de flow matching informé par graphe, résout l'imputation de données manquantes dans les systèmes spatiotemporels (qualité de l'air, trafic urbain). Il remplace le prior gaussien standard par un prior construit via filtrage spatiotemporel, et utilise un modèle de champ vectoriel hybride combinant attention spatiale, attention temporelle et propagation spatiotemporelle. Code disponible.
Lire la source
Ton avis ?
PapersBenchmarksRaisonnement

Résumé généré par Claude — vérifié par l'humain