binaryclassification

This is a project done for DS-GA 1001 on the MAGIC Telescopic Dataset

Collaborators: Jeffrey Tumminia

Goal: Build a classifier for identifying gamma rays from noisy simulated data.

Learning Objective: Compare and constrast different supervised classification algorithms in effectiveness and efficiency.

Challenges:

• Difficult classification dataset
• Low computational capacity
• Limited subject matter knowledge in feature analysis

Procedure:

DataView

1. process csv dataset, create stratified training/val/test sets

FeatureAnalysis

2. Compare relevance and information gain of features using DT, Univariate ROC and Correlation Matrix

FeatureEngineering

3. Build non-linear combinations of features and perform feature analysis on new features

Baseline Models

4. Build DT, DF, LR , SVM, Grad Boost and KNN Models as baselines for comparison

CrossVal & Parameter Tuning

5. Perform XVal tuning for DT, RF, and LR

Tuned+Voter

6. Show performace of tuned DT and LR + Build a voting classifier to improve overall classification.

See paper for more details and references: