Skip to main content
. Author manuscript; available in PMC: 2023 Jan 17.
Published in final edited form as: Neuroimage. 2022 Nov 7;264:119737. doi: 10.1016/j.neuroimage.2022.119737

Table A2.

We show how hidden dimensions of different modules of the model affect classification performance. As we do not fine-tune the hyper-parameters rigorously for each experiment, it is possible to get better results than ones reported in the main body of the paper. Similar results were seen for other datasets as well. We also show how removing the temporal attention reduces the model’s classification performance. None means the final connectivity matrix Wf was just the average of each Wt.

Dataset biLSTM dim. Self-attention dim. γ 2 Temporal Attention Mean AUC Median AUC
FBIRN 100 48 0.05 GTA 0.86 0.861
FBIRN 100 48 0.05 None 0.733 0.764
FBIRN 100 64 0.05 GTA 0.858 0.861
FBIRN 128 64 0.025 GTA 0.865 0.875
FBIRN 128 64 0.025 None 0.761 0.778
FBIRN 64 32 0.05 GTA 0.849 0.858