Dual autoencoder with reconstruction-level contrast.
<p>Left: an attributed network <i>G</i> and an augmented view produced by graph data augmentation. Middle: a shared graph-attention encoder yields node embeddings <i>Z</i>, which feed two decoders: a structure decoder reconstructing and an attribute decoder reconstructi...
Spremljeno u:
| Glavni autor: | |
|---|---|
| Daljnji autori: | , , |
| Izdano: |
2025
|
| Teme: | |
| Oznake: |
Dodaj oznaku
Bez oznaka, Budi prvi tko označuje ovaj zapis!
|
| Sažetak: | <p>Left: an attributed network <i>G</i> and an augmented view produced by graph data augmentation. Middle: a shared graph-attention encoder yields node embeddings <i>Z</i>, which feed two decoders: a structure decoder reconstructing and an attribute decoder reconstructing for the two views, yielding and . Right: reconstruction-level contrast compares, for each node <i>i</i>, the reconstructions via and ; it minimizes <i>D</i> when and , and enforces a learnable margin <i>m</i> when or . Color coding: green nodes denote non-augmented nodes; red nodes denote augmented nodes; the dotted green arc indicates minimization of <i>D</i>; the dashed orange arc indicates margin enforcement; cross-hatched bars mark augmented features; gray edges are neutral; blue heatmaps depict reconstructed matrices.</p> |
|---|