Using Machine Learning to Predict Personal Income: A Practical Project Analysis of the UCI Adult Census Income Dataset

This project is based on the UCI Adult Census Income Dataset, focusing on the binary classification problem of predicting whether an individual's annual income exceeds $50,000. It covers data exploration, feature engineering, comparison of multiple models, and optimization strategies, providing a complete technical practice reference for classification problems.

Project Background and Dataset Introduction

Income prediction is one of the classic binary classification problems in the field of machine learning. The UCI Adult Census Income Dataset is derived from U.S. census data, containing approximately 48,000 records, with the goal of predicting whether an individual's annual income exceeds $50,000. This dataset includes 14 feature variables such as age, education level, occupation, and marital status, and has typical data quality issues like missing values, class imbalance, and categorical variables that need encoding.

Exploratory Data Analysis (EDA)

Through statistical distribution visualization, it was found that people with an income exceeding $50,000 account for about 24% of the total samples, showing obvious class imbalance. Feature analysis reveals a strong correlation between education level, occupation type, and working hours and income level.

In the data preprocessing stage, missing values are handled, numerical features are standardized, and one-hot encoding is applied to categorical variables. Feature engineering also creates interaction features between years of education and occupation categories.

Model Selection and Implementation

The project implements a comparison of three mainstream classification algorithms:

Decision Tree

As a baseline model, hyperparameters such as maximum depth and minimum samples for splitting are optimized via grid search to prevent overfitting.

Random Forest

Integrates multiple decision trees, reduces variance through the Bagging strategy, and can provide feature importance ranking.

Multi-Layer Perceptron (MLP)

Constructs a neural network structure with hidden layers, adjusting learning rate, batch size, and regularization parameters to explore performance boundaries.

Model Evaluation and Comparison

Evaluation metrics include accuracy, precision, recall, F1 score, and AUC-ROC curve. The results show that Random Forest has the best overall performance, balancing accuracy and efficiency; MLP has strong expressive power but its advantages are not obvious in medium-sized datasets.

Feature importance analysis indicates that capital gains, education level, marital status, and age are key factors affecting income prediction, which is consistent with socio-economic research.

Optimization Strategies and Hyperparameter Tuning

Cross-validation is used for robust evaluation, and hyperparameter optimization uses a combination of grid search and random search. For class imbalance, SMOTE oversampling and undersampling techniques are tried; moderate balancing can improve the recognition ability of minority classes, but caution is needed against noise introduced by oversampling.

Practical Significance and Application Value

Income prediction models can be applied to financial credit risk assessment, human resource salary strategy formulation, and socio-economic factor analysis for public policies.

This project demonstrates a complete machine learning engineering practice from data cleaning to model deployment, providing a reproducible learning path for beginners and a baseline system reference framework for practitioners.