HardiVv
commented
1 month ago Configurable training features for binary classification task
The task was the same as in #62 but for binary classification.
- Surprisingly Kinematics + 13 features + lifetimes had the best outcomes.
- Kinematics + 13 features was behind by a small margin.
Some plots didn't turn out great maybe because of tensorboard?
(JP: if prediction values are very small/large in the initial epochs, tensorboard histograms are not able to visualize it correctly)
Other graphs showed the opposite, with Kinematics + 13 features performing better.
Training ROC curves were worse than validation but more stable over increasing steps. Validation ROC curves fluctuated significantly. This discrepancy is intriguing.
(JP: train being significantly worse than validation seems like a bug)
Train ROC curves
Validation ROC curves. Kinematics + features is better at low background, while Kinematics + features + lifetimes performs better overall. Choosing the right feature set is crucial depending on the model's objective.
Ran 3 binary classification tasks on three feature sets
Each run took about a day and 150 GB memory