The goal of this task is to distinguish between two types of events that occur when particles collide at high energies in the ATLAS experiment at the Large Hadron Collider (LHC) at CERN. One type of event is the decay of a Higgs boson into two tau leptons, which is a rare and important process that reveals the origin of mass. The other type of event is the background, which consists of other processes that mimic the signal but are not related to the Higgs boson. The challenge is to find a way to maximize the discovery significance of the signal over the background, using a statistical measure called the approximate median significance (AMS)(Please check the notebooks for more details).
Here, I have used a hybrid boosted bagging algorithm which we can call Boosted Extremely Randomized Trees(BXT). This model builds uses an AdaBoost algorithm with a collection of Extremely Randomized Trees(ET) as estimators. This in theory, should nullify the shorcomings of each model with AdaBoost having a tendecncy to overfit and ETs having high bias, hence reduce the bias and variance especially for a an unbalanced data set such as this. Furthermore, I am doing a comparitive study between BXT and other models such as Random Forest, AdaBoost and a Deep Neural Network with two hidden layers, to analyse how well the BXT fares against these models.
A hyperparameter search was done on all models. Higher the AMS, better is the model perfomarmance.
| Model | AMS Value |
|---|---|
| BXT | 3.49 |
| Random Forest | 3.24 |
| AdaBoost | 2.97 |
| DNN | 2.16 |
-
BXT seems to have outperformed all the models taken into consideration!
-
DNN has a tendency to overfit even through dropout layers were added and batch sizes were small. More tuning is needed
The CERN data can be download from here