ga★
Garvit AroraPORTFOLIO 2026 — AI & SOFTWARE

CREDIT CARD FRAUD DETECTION

Machine Learning / Financial Technology — Imbalance-Aware Classification on Transactional Data

BRIEF

Fraudulent transactions are uncommon compared with legitimate transactions, making classification challenging and rendering naive accuracy metrics misleading.

ENGINEERING IDEA★

Prepare transaction data, address class imbalance with SMOTE, train ensemble classification models, and evaluate performance with precision, recall, F1-score, ROC-AUC, and confusion matrices.

TOOLS USED

Python · Scikit-learn · Random Forest · SMOTE · Pandas · NumPy

PROJECT TYPE

*Machine Learning / Financial Technology
*Machine Learning Experiment
01 // SYSTEM ARCHITECTURE & VISUAL SCHEMATIC
PIPELINE // SMOTE + RANDOM FOREST CLASSIFIERMETRICS: PRECISION · RECALL · F1 · ROC-AUC
FEATURE SPACE PARTITION // SMOTE MINORITY RESAMPLING● MAJORITY (LEGITIMATE) ● FLAGGED ANOMALY ○ SMOTE SYNTHETIC SAMPLEEVALUATION FRAMEWORKTRUE POS (TP)Recall FocusFALSE POS (FP)Precision CostFALSE NEG (FN)Critical RiskTRUE NEG (TN)SpecificityPRIMARY METRICS:ROC-AUC · F1-SCORE · PR-CURVE
02 // 4-STAGE WORKFLOW PROGRESSION
STAGE 01

Exploratory Data Analysis & Scaling

Inspect feature distributions, normalize transaction amount and time variables, and isolate class skew.

STAGE 02

SMOTE Minority Resampling

Synthesize balanced training distributions in feature space to prevent majority-class dominance.

STAGE 03

Random Forest & Model Comparison

Fit ensemble tree classifiers and benchmark against baseline classification approaches.

STAGE 04

Imbalance-Aware Evaluation

Analyze Confusion Matrices, Precision-Recall balance, F1-score, and ROC-AUC curves.

03 // IMPLEMENTATION, OUTCOMES & KEY LEARNINGS
IMPLEMENTATION DETAILS
  • ★Executed data preprocessing and feature scaling pipelines using Pandas, NumPy, and Scikit-learn.
  • ★Integrated SMOTE from imbalanced-learn within the training workflow to mitigate severe majority-class bias.
  • ★Trained Random Forest classifiers and compared decision behavior across resampled and baseline splits.
  • ★Structured evaluation around confusion matrices, precision, recall, F1-score, and ROC-AUC to reflect the operational cost of false negatives versus false positives.
VERIFIED OUTCOMES
  • ✓Produced a complete, reproducible machine learning experiment demonstrating imbalance-aware classification methodology.
  • ✓Highlighted why recall, precision, F1-score, and ROC-AUC are essential when evaluating rare-event detection models.
REFLECTIONS & NEXT STEPS
  • →Resampling must be strictly confined to training partitions to prevent data leakage into validation and test sets.
  • →Future work includes cost-sensitive threshold tuning and gradient-boosted trees with calibrated probability outputs.