Income Prediction Based on 1994 Census Data: A Classic Case Study of Machine Learning Classification Problems

This project is based on the 1994 U.S. Census Adult dataset, focusing on the binary classification problem of "whether an individual's annual income exceeds $50,000". It covers the complete machine learning workflow including data exploration, feature engineering, model training, and evaluation. By comparing multiple models such as logistic regression and random forest, it provides learners with practical references for real data processing and classification tasks, making it a classic case for machine learning beginners.

Income prediction is of great value in fields such as policy-making, credit evaluation, and marketing. The Adult dataset (Census Income dataset) used in this project comes from the U.S. Census Bureau in 1994, containing 48842 records. Each record includes 14 input features (demographic + employment-related) such as age, education level, and occupation, as well as the target variable "whether income > $50K/year". The advantages of this dataset are: moderate size, diverse feature types (numeric + categorical), real-world issues like missing values and class imbalance, making it suitable for practicing complete data processing skills.

Raw data needs to be processed through the following steps: 1. Missing value handling: For missing values in fields like Workclass and Occupation, filling strategies (mode/median) are used to maintain integrity; 2. Categorical encoding: Unordered categories (e.g., Race, Sex) use one-hot encoding, while ordered categories (e.g., Education) use label encoding; 3. Numeric scaling: Standardization or normalization is applied for linear models/neural networks; 4. Feature selection: Redundant features are removed through correlation analysis and model importance evaluation (e.g., Education and Education-Num are highly correlated, so one is retained).

EDA reveals: 1. Imbalanced target variable: 76% of samples have income ≤ $50K, 24% > $50K; 2. Univariate distribution: Age is concentrated between 20-50 years old, education years are mostly at high school level, and working hours are mainly 40 hours; 3. Bivariate relationships: Higher education and executive management occupations are strongly correlated with high income; 1994 data shows that the proportion of high income among males is higher than females; 4. Multivariate: Education and Education-Num are highly correlated, so collinearity should be noted.

The project implements multiple classification models: logistic regression (baseline model), decision tree, random forest, gradient boosting tree, and SVM. For evaluation: due to class imbalance, metrics such as precision, recall, F1 score, and ROC-AUC are used; K-fold cross-validation is applied to ensure generalization ability. Among them, random forest and gradient boosting tree perform better, effectively handling feature interactions and overfitting issues.

Application Scenarios: Credit evaluation (repayment ability prediction), marketing (high-value customer identification), policy research (analysis of income influencing factors), education planning (curriculum optimization). Limitations: Insufficient data timeliness (1994), lack of modern features (e.g., skill certificates, geographic location), fairness risks (gender/race features may lead to bias). Improvement Suggestions: Use updated data sources, try deep learning models, conduct fairness audits, and add combined feature engineering.