L2,1-norm-based Sparse Dynamic Graph Convolution Neural Network
Poster presentation
Research approach
Applied row-wise L2,1 sparsity to a DGCNN that learns EEG electrode connections. Classified three emotions using DE features from five frequency bands and compared L1 regularization.
View full-size figure
Results
Subject-dependent evaluation on SEED, with 15 participants: L2,1 accuracy 75.52 ± 8.88%, L1 75.56 ± 8.89% and DGCNN 74.90 ± 8.20%. Classification performance was comparable to L1.
| Model | Accuracy |
|---|---|
| DGCNN | 74.90 ± 8.20 |
| L1 Sparse DGCNN | 75.56 ± 8.89 |
| L2,1 Sparse DGCNN | 75.52 ± 8.88 |