Some Important Notes
- Every entry in the dataset had a maximum multiplicity of particles - this maximum multiplicity differed which caused me problems while flattening. Hence, I split up each particle individually into an entry and particles of each event were alloted a common ID.
- I first tried a random forest classification model which had a fairly low accuracy.
- Then moved to a Gradient Boosting LightGBM model.
- In the LightGBM model, during initial training, data analysis showed that the dataset was significantly unbalanced. Hence, to get a comprehensive view of the picture, I concatenated three different datasets, then further balanced the dataset using SMOTE.
- Due to lack of computational power, I had to cut down the dataset to 1/3rd random entries of the above.
- Then, I trained a random LightGBM model on it.
- Tried Hyperparameter tuning using Bayesian optimisation, which was a learning opportunity albeit fractional improvements in accuracy.