A Research-Oriented Portfolio of Applied Predictive Modeling

Welcome to my centralized repository of machine learning classification projects. This collection highlights end-to-end workflows spanning:

• Exploratory Data Analysis (EDA)
• Feature engineering & preprocessing pipelines
• Model development & comparison
• Hyperparameter tuning
• Evaluation & interpretation

Each project folder is self-contained, reproducible, and written in a clear, research-oriented style suitable for academic, portfolio, and professional use.

Each subdirectory typically includes:

• Jupyter Notebooks – full workflow (EDA → preprocessing → modeling → evaluation)
• Saved Models – pipelines, tuned models (.joblib, .pkl)
• Dataset – or reference links for external datasets
• Documentation – project-specific README.md

Below are 3 of the most complete, research-oriented projects in this repository. Each showcases advanced EDA, custom preprocessing, and model experimentation.

Goal: Predict whether a person earns >50K using demographic & socioeconomic features. Dataset: UCI Adult Census Income (32k rows, mixed categorical & numeric)

• Detailed EDA (imbalance, missingness patterns, heavy skew in capital gain/loss)
• Dual preprocessing tracks for tree models vs linear models
• Outlier handling using IQR & Z-score rules
• Trained 10+ models (Logistic Regression, SVC, LightGBM, CatBoost, RF, KNN, etc.)
• Saved modular pipelines for each model

• F1: 0.6253
• Accuracy: 0.8397
• Excellent performance on imbalanced data

📂 Directory: Classification/adultCensusIncome/

Goal: Detect fraudulent transactions using the massive IEEE-CIS dataset (~1M rows). Challenge: Heavy imbalance, high dimensionality, complex categoricals.

• Robust EDA with Spearman correlation (sampled due to dataset size)

• Feature reduction using correlation threshold (|ρ| > 0.9)

• Skew-based feature grouping → custom preprocessing pipelines:

  • Standard scaling
  • Robust scaling
  • Power transform (Yeo-Johnson)

• Advanced categorical encoders (Ordinal vs Target Encoding based on model family)

• Trained multiple models (RF, XGBoost, AdaBoost, GBoost, NB, KNN, etc.)

• Accuracy: 0.9817
• F1 Score: 0.6642
• ROC-AUC: 0.9468

Includes learning curve & feature importance analysis.

📂 Directory: Classification/IEE-Fraud-Detection/

Goal: Classify tumors as benign or malignant. A well-structured project with clean pipeline and strong evaluation.

• Preprocessing pipeline with StandardScaler + imputation
• Multiple models trained & compared
• Balanced dataset → focus on precision/recall tradeoffs
• Visualizations: correlation matrix, distributions, pairplots, classifier curves

• Accuracy: 0.9649
• Precision: 0.9857
• Recall: 0.9583
• F1 Score: 0.9718

📂 Directory: Classification/breastCancerWisconsin/

These projects are smaller or exploratory, included for completeness:

Predict term deposit subscription. Includes SMOTE-based balancing, model comparison, and SHAP interpretation.

Classic ML task with feature engineering (family size, title extraction).

Baseline pipeline complete → tuning pending.

EDA done → modeling next.

Combined pipeline for regression → classification setup optional.

Languages:

• Python

Core Libraries:

• pandas, numpy, matplotlib, seaborn
• scikit-learn, xgboost, lightgbm, catboost
• imblearn (SMOTE)
• scipy, statsmodels
• joblib for model persistence

Notebooks: Jupyter / Kaggle / Colab

Each project contains:

A research-style README Notebooks documenting full ML workflow Saved models Visualizations & metrics

Start by exploring any project directory listed above.