Implement new realistic data sets on the original DGI model.
Dataset: citeseer and cora
Original paper: "Deep Graph Infomax" (https://arxiv.org/abs/1809.10341)
Research question: different readout functions
Training result:
-
ML4Grpahs_DGI_cora_diffpool_100
Average accuracy: tensor([0.8184]) accs mean: tensor(81.8380) accs std: tensor(0.1640)
-
ML4Grpahs_DGI_cora_average_100
Average accuracy: tensor([0.8197]) accs mean: tensor(81.9720) accs std: tensor(0.1679)
-
ML4Grpahs_DGI_cora_sum_norm_100
Average accuracy: tensor([0.8135]) accs mean: tensor(81.3460) accs std: tensor(0.1182)
-
ML4Grpahs_DGI_citeseer_diffpool_100
Average accuracy: tensor([0.7207]) accs mean: tensor(72.0680) accs std: tensor(0.2123)
-
ML4Grpahs_DGI_citeseer_average_100
Average accuracy: tensor([0.7294]) accs mean: tensor(72.9440) accs std: tensor(0.1402)
-
ML4Grpahs_DGI_citeseer_sum_norm_100
Average accuracy: tensor([0.7232]) accs mean: tensor(72.3160) accs std: tensor(0.1517)