model optimization

[BioLaoXu]

According to literature research,GraphSmote is probably the only one toolkit that can train graph neural networks on unbalanced data,It's a great privilege to use this toolkit. My learning task is consistent with the Cora data set,just a simple node-level classification(0 or 1,0 represent negative and 1 represent positive),i have ~13000 nodes(positive:~550,the rest nodes are all negative),each node has 350 features;then i rewrite the Cora data_loader function for my data set,and it words successfully.then I train the model by GraphSmote main.py： nohup python3 -u main.py --imbalance --no-cuda --dataset=cora --setting='embed_up' --epochs=1000 --model=GAT --up_scale=0.95 --nhid=8 --batch_size=64 --lr=0.001 --dropout=0.5 & It works!but the accuracy of the model in both the training data and the verification data set is very low and fluctuates within 0.6,just like bellow: Loading cora dataset... edges between class 0 and class 0: 63682080.000000 edges between class 0 and class 1: 2235402.000000 edges between class 1 and class 0: 2235402.000000 edges between class 1 and class 1: 99624.000000 0-th class sample number: 12353 1-th class sample number: 500 valid current auc-roc score: 0.643200, current macro_F score: 0.333333 Epoch: 00001 loss_train: 0.7087 loss_rec: 0.7087 acc_train: 0.5000 loss_val: 0.7085 acc_val: 0.5000 time: 125.3990s Test set results: loss= 0.7085 accuracy= 0.5000 test current auc-roc score: 0.452893, current macro_F score: 0.333333 valid current auc-roc score: 0.459200, current macro_F score: 0.333333 Epoch: 00002 loss_train: 0.7085 loss_rec: 0.7085 acc_train: 0.5000 loss_val: 0.7087 acc_val: 0.5000 time: 125.1620s valid current auc-roc score: 0.481600, current macro_F score: 0.333333 Epoch: 00003 loss_train: 0.7083 loss_rec: 0.7083 acc_train: 0.5000 loss_val: 0.7083 acc_val: 0.5000 time: 124.7956s valid current auc-roc score: 0.561600, current macro_F score: 0.333333 ...... Epoch: 00820 loss_train: 0.6924 loss_rec: 0.6924 acc_train: 0.5500 loss_val: 0.6953 acc_val: 0.3400 time: 174.9777s the detailed results: nohup.txt

i have try reset the nhid/dropout/learn ratio ,but both the acc_train and acc_val are very low and fluctuate,Do you have any suggestions for optimization,Looking forward to your reply!thanks

[TianxiangZhao]

How do you set the im_ratio? You can try set it as 0.5 and test the performance.

[neeraja1504]

I am working on a similar problem involving GATs. However, when I keep nhid as a default 64 there is a tcmalloc error. The GAT model is also taking a lot of hours to train compared to the GCN and GraphSage models. The accuracy is around 55% even when trained for more than a thousand epochs. Any thoughts on what can be done to improve the performance and reduce the training time?