I am trying to perform a regression task using SAGEConv

Models & Apps
#1

For that I am having an excel sheet for regressors and a target value. A cyclic graph is created between each of the sample and features will be the attributes of nodes. The SAGEConv last layer is a Linear layer with one node. Loss Function is MSE. But this is performing very poor, Can any one suggest some good technique for regression using GNN?

#2

It seems that your dataset is a single table. Why are you building a cyclic graph out of it? In the worst case, the rows of the table might be independent to each other, and running GraphSAGE on it will probably just smooth out the predictive signal on the row itself and lead to worse performance. Have you compared it with simple baselines such as MLP, XGBoost or AutoGluon?

1 Like
#3

Thanks for reply I want to compare the results with GNN only.
The rows are time stamps so cyclic connection is the only way possible.
The code is working though with poor performance.
And the results are better when I removed linear layer and make the output sage layer with out_feat=1

I am wondering where in dgl library the depth factor K is defined for SAGEConv? Is K and Layer same, I dont think so, if they are not what is the default value of K?