Poor Prediction Results

[oliverhvidsten]

Multi-Task model with 9,100 points for one property and 6,200 points for the other property I have previously trained this dataset on polygnn_trainer and got much better results

I have attached:

• three parity plots (plots.png),
• the output log file (mT_Log.txt)
• the altered python file that I used to run polygnn (GNN_CV_training.txt)

Do you know why this behavior might be seen? I got very similar results on my first test and decided to increase some of the constants at the beginning. Those changes did not improve prediction accuracy.

mT_Log.txt GNN_CV_training.txt

[rishigurnani]

@oliverhvidsten does this problem persist on smaller data sets (e.g., 100 data points per property)? If so, would you be willing to share the smaller data set with me so I can reproduce and investigate?

[oliverhvidsten]

After doing some testing, it appears that the information in graph_feats might not be accounted for while the model is training. The predictions are invariant across all of the non-chemical features.

[rishigurnani]

Wow, you are totally right. There's a bug in the code. Above I linked a PR. Can you test it to see if it solves your issue?

[oliverhvidsten]

Looks a lot better after the fix!