基于Ames数据集的房价预测——模型比较与超参数优化

House Price Prediction Based on the Ames Dataset—Model Comparison and Hyperparameter Optimization

摘要: 房价受到多个因素的影响，对其的预测存在非线性、高维度及数据异方差性问题，本文基于Kaggle House Prices数据集，构建了一套包含线性模型、集成学习与核方法的多元回归预测框架。本研究选取支持向量回归(SVR)、随机森林(RandomForest)、XGBoost、LightGBM及弹性网络(ElasticNet)五种模型对房价进行预测。实验对每种模型采用5折交叉验证，以对数均方根误差(log-RMSE)、对数平均绝对误差(log-MAE)及决定系数(R²)作为评价指标。研究结果表明这五种模型中最优的是SVR，三个评价指标都是最优的。基于前面五种模型的对比，选取SVR模型进行调优，通过Optuna贝叶斯优化对SVR中的三个超参数进行调优后，模型性能进一步提升，log-RMSE降至0.1249，R²为0.8985，显著优于其他对比模型。

Abstract: House Prices are influenced by multiple factors, and there are problems of nonlinearity, high dimensionality and data heteroscedasticity in their prediction. Based on the Kaggle House Prices dataset, this paper constructs a multiple regression prediction framework including linear models, ensemble learning and kernel methods. This study selects five models, namely Support Vector Regression (SVR), RandomForest, XGBoost, LightGBM and ElasticNet, to predict housing prices. For each model in the experiment, 5-fold cross-validation was adopted, and the logarithmic root mean square error (log-RMSE), logarithmic mean absolute error (log-MAE), and coefficient of determination (R²) were used as evaluation indicators. The research results show that among these five models, SVR is the best, and all three evaluation indicators are optimal. Based on the comparison of the previous five models, the SVR model was selected for tuning. After tuning the three hyperparameters in SVR through Optuna Bayesian optimization, the model performance was further improved. log-RMSE dropped to 0.1249 and R² was 0.8985, which was significantly better than other comparison models.