search

divelab / lgcn

GNU General Public License v3.0
136 stars 65 forks source link

some question about the select adjacent nodes features？ #7

Closed xuesanshi closed 5 years ago

xuesanshi commented 5 years ago 
    First of all, thank you for answering my question many times，When I read your thesis, I found that you used the maximum pooling features of K nodes when you selected the neighborhood information of a point. But can I consider learning a weighted adjacency matrix first, and get the neighborhood features according to the weight of the center point and the neighboring point?think you
HongyangGao commented 5 years ago 
    First of all, thank you for answering my question many times，When I read your thesis, I found that you used the maximum pooling features of K nodes when you selected the neighborhood information of a point. But can I consider learning a weighted adjacency matrix first, and get the neighborhood features according to the weight of the center point and the neighboring point?think you

Yes, this is used to select fixed number (k) of nodes for the center node to do convolution. However, this selection is performed in each channel separately. This means it may select different nodes for different channels. Hope this helps.