GRAPHMAE · A VISUAL FIELD GUIDE Learning from Hide a feature row. Let the graph provide context. Then ask what a good reconstruction actually means.
Start reading → B is withheld, but its neighbors remain Nodes A to F; edges AB, BC, BD, DE, EF, CF. Masked feature rows: B. Dashed circles denote masked rows; all nodes and edges remain. A (2, 1) B MASK C (2, 2) D (3, 1) E (1, 3) F (2, 3) B's original row is withheld. A, C and D remain connected to it. Synthetic graph, dimensionless features.
Five chapters for an applied-ML reader. Message passing, notation and cosine geometry are refreshed when needed. Trace the original method, test a prescribed loss comparison, then design an honest transfer evaluation.
01 A graph gives context A missing row is still attached to its neighbors. That is the opportunity—and the constraint.
↗ 02 Hide a row. Keep its target. Follow the original GraphMAE computation, including the second mask before the decoder.
↗ 03 What survives the mask? Mask placement changes available context. A fixed reconstruction table does not measure how hard a model would find it.
↗ 04 Direction is a choice Scaled cosine error rewards aligned features. Squared error also sees their magnitude.
↗ 05 Beyond this graph A reusable encoder is a starting point. Transfer is an evaluation claim.
↗