Predicting China's Marriage Rate using Machine Learning Models and Cross-Validation Evaluation

After China's accession to the WTO, the country has experienced rapid development over the past 20 years, but the marriage rate has been declining. This study aims to analyze marriage data using machine learning, identify the factors influencing the marriage rate, and make predictions. The data covers seven variables including GDP, housing prices, fertility rate, and education level in 31 regions of China from 2003 to 2022. The study employed five machine learning algorithms with five models: Ridge, Lasso, Elastic Net, Polynomial, and Weighted Least Squares, along with three types of Cross-Validation techniques – Holdout, LOOCV, and K-Fold. By comparing the five regression models and evaluating the marriage rate models through Holdout, LOOCV, and K-Fold Cross-Validation, the results show that the Polynomial regression exhibits higher predictive accuracy, with MSE of 3.473 for Holdout, 3.001 for LOOCV, and 3.018 for K-Fold. Following closely is Ridge regression with MSE of 3.565 for Holdout, 3.326 for LOOCV, and 3.334 for K-Fold. These two models outperform the less performing Lasso and Elastic Net regressions. Weighted Least Squares show stable but slightly inferior performance. By comparing the predicted values for 2022 with the actual values, it is confirmed that Polynomial and Ridge regression closely align with the actual values, predicting the trend of changes in real and predicted values for each province, highlighting their effectiveness in predictive tasks involving complex data patterns.