Census Income Classification

📌 Project Overview

This project follows the complete machine learning life cycle to solve a real-world classification problem using the 1994 Census Income Dataset. The objective is to predict whether an individual's income exceeds $50K/year based on demographic and work-related attributes.

This notebook includes:

• Exploratory data analysis (EDA)
• Data cleaning and preprocessing
• Feature engineering and selection
• Model training and evaluation using:
  • Logistic Regression
  • Random Forest
  • Neural Networks

---

📁 Dataset Used

File: censusData.csv
Source: UCI Machine Learning Repository (modified for educational use)

🎯 Problem Definition

• Type: Supervised Learning
• Category: Binary Classification
• Target Variable: income_binary (0 = ≤50K, 1 = >50K)
• Features:
  • Continuous: age, fnlwgt, education-num, capital-gain, capital-loss, hours-per-week
  • Categorical (one-hot encoded): workclass, education, marital-status, occupation, relationship, race, sex_selfID, native-country

---

⚙️ ML Workflow

🧼 Data Preprocessing

• Imputed missing values using column means (for age and hours-per-week) or dropped incomplete rows
• Encoded the target variable using LabelEncoder
• Applied one-hot encoding to all categorical features
• Standardized features using StandardScaler

🔍 Feature Selection

• Used Random Forest to extract the top 50 most important features to reduce noise

---

🤖 Models Implemented

1. Neural Network (TensorFlow/Keras)

• 3 hidden layers with ReLU activation
• Batch Normalization and Dropout used for regularization
• Optimizers: Tested with both SGD and Adam
• Trained for 115 epochs with early insight logging every 5 epochs

Performance:

• Accuracy on test set: ~83.9%
• Good generalization with training vs. validation accuracy curves closely aligned

2. Logistic Regression (Baseline Model)

• Implemented with sklearn.linear_model.LogisticRegression
• Achieved comparable performance to neural network

Performance:

• Accuracy on test set: ~84.2%
• Simple, fast to train, and interpretable

---

📊 Evaluation

• Metric used: Accuracy
• Also evaluated predicted probabilities vs. actual classes
• Plotted training/validation loss and accuracy over epochs
• Compared model complexity vs. performance for fairness

---

✅ Key Takeaways

• Model Selection matters: simple models like Logistic Regression can match the performance of more complex ones for some datasets.
• Feature Engineering and Preprocessing are critical: handling missing data and one-hot encoding were essential to model success.
• Bias/Fairness Consideration: This problem sheds light on how income inequality correlates with social and demographic features, which can have ethical implications in real-world deployments.

---

🛠️ Technologies Used

• Python 3
• Pandas, NumPy
• Scikit-learn
• TensorFlow / Keras
• Matplotlib, Seaborn

---

📚 Author Notes

This notebook was completed as Lab 8 of the Machine Learning curriculum. The lab required defining a predictive task, developing a complete ML pipeline, and reflecting on model performance and fairness. The Census dataset was chosen for its rich demographic insights and relevance to social equity research.