search

Ljiajie / Addgraph

The code of Addgraph frame.
20 stars 7 forks source link

What's the difference between _U_ ahd _D_ #2

Open Sebastianyang666 opened 1 year ago

Sebastianyang666 commented 1 year ago

Hello.what's the difference between UCI_U_ADDgraph and UCI_D_ADDgraph？Which one is the eassy model ? I didn't find Attention in UCI_U_ADDgraph or GCN in UCI_D_ADDgraph

Ljiajie commented 1 year ago

Hello.what's the difference between UCI_U_ADDgraph and UCI_D_ADDgraph？Which one is the eassy model ? I didn't find Attention in UCI_U_ADDgraph or GCN in UCI_D_ADDgraph

Hi! The sign 'U' in UCI_U_ADDgraph means undirected graphs and 'D' in UCI_D_ADDgraph means directed graphs. The core difference between GAT and GCN is how to extract the feature representation of the neighbor node. GAT replaces the fixed, standardized operations of GCN with an attention mechanism. Essentially, GAT simply replaces the normalized functions of the original GCN with neighbor node feature aggregation functions using attention weights. However, it should be noted that the Laplacian matrix in the theoretical derivation of GCN is symmetric, while for the directed graph, the Laplacian matrix is asymmetric and cannot be eigenseparable(more details can refer to the original papers of GAN and GAT). Therefore, for undirected graphs we use GCN modeling and GAT for directed graphs. Of course, there are many improved ways to adapt GCN to directed graphs at present.

Sebastianyang666 commented 1 year ago

Hello.what's the difference between UCI_U_ADDgraph and UCI_D_ADDgraph？Which one is the eassy model ? I didn't find Attention in UCI_U_ADDgraph or GCN in UCI_D_ADDgraph

Hi! The sign 'U' in UCI_U_ADDgraph means undirected graphs and 'D' in UCI_D_ADDgraph means directed graphs. The core difference between GAT and GCN is how to extract the feature representation of the neighbor node. GAT replaces the fixed, standardized operations of GCN with an attention mechanism. Essentially, GAT simply replaces the normalized functions of the original GCN with neighbor node feature aggregation functions using attention weights. However, it should be noted that the Laplacian matrix in the theoretical derivation of GCN is symmetric, while for the directed graph, the Laplacian matrix is asymmetric and cannot be eigenseparable(more details can refer to the original papers of GAN and GAT). Therefore, for undirected graphs we use GCN modeling and GAT for directed graphs. Of course, there are many improved ways to adapt GCN to directed graphs at present.

I totally understand now,thank you so much for you detail reply.