🏦 Bank Loan Approval - Binary Classifier

A Sequential classification model trained on a large real-world bank loan dataset from Kaggle to predict loan approval.

📊 Dataset

Customer financial and loan data 📥 Dataset for Bank Loan Prediction

⚙️ Workflow

• EDA & Cleaning: Handled nulls, outliers, patterns
• Preprocessing: Encoded categoricals, scaled features
• Model: Sequential: NN with 32,768 parameters, ReLU layers, sigmoid output
• Training: Binary cross-entropy + Adam + early stopping

🧠 Features used

• Loan Status: Indicates whether the loan was approved or denied (target variable for classification).
• Current Loan Amount: The total amount the client is requesting or currently owes on the loan.
• Term**: The length of the loan repayment period**, usually in months (e.g., 36 or 60 months).
• Credit Score: A numerical value showing the client’s creditworthiness based on past credit behavior.
• Annual Income: The client's yearly earnings, used to assess their ability to repay the loan.
• Years in Current Job: The duration of employment at the current job, indicating job stability.
• Home Ownership: Shows if the client rents, owns, or has a mortgage, reflecting financial stability.
• Purpose: The reason for taking the loan (e.g., debt consolidation, medical, education, etc.).
• Monthly Debt: The total monthly payments for debts (excluding the new loan), used in debt-to-income ratio.
• Number of Open Accounts: Total number of active credit accounts (like loans or credit cards) the client currently has.
• Current Credit Balance The outstanding amount the client owes across all credit accounts.
• Bankruptcies Number of times the client has declared bankruptcy, signaling past financial distress.
• Tax Liens: Number of legal claims by the government on the client’s property due to unpaid taxes.

📁 Files

• LoanPrediction.ipynb – Full pipeline
• Kaggle Notebook - Full Explanation

🚀 Results

Achieved strong classification performance of 80% on unseen test data with a well-generalized and compact model.

✈️ Airline Delay & Cancellation Prediction - Binary Classifier

A Sequential classification model trained on a large real-world airline dataset to predict whether a flight will be delayed by 15+ minutes or canceled.

📊 Dataset

Historical flight data enriched with weather, airport, and carrier features.
2019 airline delays and cancellations
🎯 Target: DEP_DEL15 – Binary indicator for delay >15 minutes or cancellation.

⚙️ Workflow

• EDA & Cleaning: Handled missing values, outliers, anomalies
• Preprocessing:
  • Encoded categorical features
  • Applied feature scaling
  • Log-transformed skewed features
• Custom Class Weights: Balanced training for imbalanced label distribution
• Model:
  • Deep Neural Network with multiple Dense layers, Dropout, and BatchNormalization
  • Total parameters: ~35K
  • Loss: Binary cross-entropy
  • Optimizer: Adam
  • Metrics: Accuracy, Precision, Recall, AUC

🧠 Features Used

• MONTH: Month of flight
• DAY_OF_WEEK: Day of the week
• DEP_TIME_BLK: Departure time block (e.g., 0600–0659)
• DISTANCE_GROUP: Distance bucket of the flight
• SEGMENT_NUMBER: Flight segment order
• CONCURRENT_FLIGHTS: Concurrent departing flights in same time block
• NUMBER_OF_SEATS: Seats available on aircraft
• CARRIER_NAME: Airline carrier
• AIRPORT_FLIGHTS_MONTH: Avg. airport traffic
• AIRLINE_FLIGHTS_MONTH: Avg. flights operated by airline
• AIRLINE_AIRPORT_FLIGHTS_MONTH: Combined airport-airline traffic
• PLANE_AGE: Age of aircraft
• LATITUDE, LONGITUDE: Geolocation of airport
• PRCP, SNOW, SNWD, TMAX, AWND: Weather conditions (precipitation, snow, temp, wind)

📁 Files

• LoanPrediction.ipynb – Full pipeline

• Kaggle Notebook - Full Explanation

🚀 Results

Achieved ~65% accuracy on unseen test data using a custom-weighted loss function and tuned architecture.
The model shows strong generalization for real-world flight scheduling and operations forecasting.

🏡 House Price Prediction - Regression Model

A machine learning regression model trained on the well-known California housing dataset to predict house prices based on key demographic and geographic features.

📊 Dataset

California housing dataset from sklearn.datasets, derived from the 1990 U.S. Census.
📥 California Housing Dataset Documentation

⚙️ Workflow

• EDA & Cleaning: Inspected and visualized data distribution, identified and handled outliers
• Preprocessing: Scaled features using StandardScaler
• Model: Trained a Linear Regression model and a Tree-based regressor (e.g., RandomForestRegressor)
• Training: Used mean_squared_error as the loss metric and evaluated performance using MAE, RMSE, and R² score

🧠 Features Used

• MedInc: Median income in the block group (normalized)
• HouseAge: Median age of houses in the area
• AveRooms: Average number of rooms per household
• AveBedrms: Average number of bedrooms per household
• Population: Total population in the block group
• AveOccup: Average number of household members
• Latitude: Geographical latitude
• Longitude: Geographical longitude
• 🎯 Price: Median house value (target variable, scaled)

📁 Files

• HousePricePrediction.ipynb – Full code pipeline with training, evaluation, and visualizations
• Colab Notebook – Explained end-to-end

🚀 Results

Achieved strong regression performance on test data:

• MAE: ~0.1582
• RMSE: ~0.2072
• R² Score: ~0.82
  Demonstrated good generalization and accurate predictions with well-preprocessed features and effective model choice.

🔐 Fraud Detection Using Machine Learning & Deep Learning

CreditCardFraudDetection.ipynb
This project aims to detect fraudulent transactions using both Machine Learning (Random Forest) and Deep Learning (Neural Networks) approaches. The focus is on classifying transactions as fraudulent or genuine using six key financial features.

• live Colab workFlow – Full pipeline