search

SIMEXP / fmri_predict

predicting fmri activaties from connectome
15 stars 11 forks source link

Test GCN with different brain graphs #21

Open zhangyu2ustc opened 5 years ago

zhangyu2ustc commented 5 years ago

the first step of GCN is to construct a populational graph. there are three types of brain graphs we could choose:

  1. spatial neighborhood
  2. structural covariance
  3. RSFC

In the 2nd and 3rd graphs, we need to additionally construct a sparse KNN graph because it is very dense which might cause computational problems

Note, previously I had made a mistake by taking the weakest connections for the KNN graph, which has worsen the performance of GCN. Now it worked and we could compare them.

zhangyu2ustc commented 5 years ago

the performance of using different brain graphs are very similar. for motor classification (epochs=40): train_acc test_acc spatial: 0.94 0.90 SC: 0.93 0.90 RSFC: 0.93 0.90

we need to figure out the reason. After plotting the Adj-matrix using imshow, we do see a big difference between brain graphs. Why the performance did not change? maybe during KNN-graph setting?