Closed lijiaoyang closed 1 year ago
This is a known limitation for GCNs in general. However, there are a few directions to mitigate this issue. The one mentioned in the paper is by creating a random subset of unlabelled < entire unlabelled pool. Another way is to sub-graph the feature space according to groups of your labelled data. (This one I also had it implemented and it works). Finally, you can look into graph coarsening. There are some more elegant solutions in there.
thanks for your reply!! about "Another way is to sub-graph the feature space according to groups of your labelled data. ", could you be more specific? Or any example code?
thanks again !!
For large data sets such as Imagenet, when constructing adjacency matrix, the program memory and video memory will explode. Is there any good solution