Analysis of the Telescope Gamma Dataset

Overview of the repository

• ./report.pdf : the report of the analysis
• ./visualization.ipynb: it contains the code for the analysis of the dataset
• ./processing_model.ipynb: it contains the code for running the experiments and the two best tuned model pipelines
• ./curves.ipynb: it contains the code for the creation of the learning curves and the roc curves
• ./images/* : all the images used for the report
• ./images/outliers_experiments/* : the confusion matrices and roc curves of the experiments for each model
• ./images/outliers_experiments/results_experiment.pdf : The experiment results in a compact format

Dataset

The data present in this dataset are generated to simulate registration of high energy gamma particles in a ground-basedatmospheric.Cherenkov gamma telescope (CTA) observes high energy gamma rays, taking advantage of the radiation emitted by chargedparticles produced inside the electro-magnetic showers.

Depending on the energy of the primary gamma, a total of few hundreds to some 10000 Cherenkov photons get collected, inpatterns (called the shower image), allowing to discriminate statistically those caused by primary gammas (signal) from theimages of hadronic showers initiated by cosmic rays in the upper atmosphere (background).

The goal of the classification is to correctly classify the gamma signal (g) from hadron (h, background).

The dataset is composed of 19020 distinct records with 11 attributes and does not contain missing values.

Attribute	Type	Description
fLength	Real	major axis of ellipse
fWidth	Real	minor axis of ellipse
fSize	Real	10-log of sum of content of all pixels
fConc	Real	ratio of sum of two highest pixels over fSize
fConc1	Real	ratio of highest pixel over fSize
fAsym	Real	distance from highest pixel to center, projected onto major axis
fM3Long	Real	3rd root of third moment along major axis
fM3Trans	Real	3rd root of third moment along minor axis
fAlpha	Real	angle of major axis with vector to origin
fDist	Real	distance from origin to center of ellipse
class	Categorical	g gamma (signal), h hadron (background)

The Dataset is clearly imbalance since there are 12332 (65%) samples labelled with Gamma and 6688 (35%) labelled as Hadron.

Training Pipeline (more details in ./report.pdf)

1. Local Outlier Factor
2. Stratified Training/Test split
3. Robust Scaler
4. SMOTE / BorderLineSMOTE
5. Stratified Cross Validation
6. Random Grid Search

Experiments (more details in ./report.pdf)

Multiple experiments have been performed to truly understand the importance of each step in the training pipeline. More precisely, the analyzed steps are the Outliers Removal, the application of the Scaler and the type of Up-sampling.

Ablation	Parameter
LocalOutlierFactor	[2, 3,5,10,15,20,False]
Scaler	[True, False]
Up-Sampling	["BorderLineSMOTE", "SMOTE"]

To running the experiments, a simple nested loops is used in order to cover all the possibilities and then, for each model, is used a random grid search.