Author(s)

Xiaoyang Li

Corresponding Author

Xiaoyang Li

Abstract

This study focuses on optimizing the prediction of Shanghai's second-hand housing prices using a Convolutional Neural Network (CNN) model. Traditional second-hand housing price prediction methods have limitations in handling high-dimensional data and mining spatial correlations, leading to suboptimal accuracy. To address this, we collected 2024 Shanghai second-hand housing data from Kaggle, involving 171048 samples and nearly 30 features, which were preprocessed (including cleaning, encoding, and dimensionality reduction). A 1D-CNN model was constructed, and optimization strategies were implemented: deepening the network structure (adding convolutional and fully connected layers), strengthening regularization (incorporating L2 regularization and Dropout), and testing combinations of activation functions (ReLU, LeakyReLU, ELU) and optimizers (Adam, RMSprop). Experimental results show that the optimized model (ELU-Adam combination) outperforms the original CNN model, with MSE reduced by 59.0% (to 8362.02), MAE reduced by 42.7% (to 54.46), and R² improved to 0.9546, indicating stronger predictive accuracy and generalization ability. This research provides a reliable decision-making basis for stakeholders in the real estate market.

Keywords

Second-Hand Housing Price Forecasting, Convolutional Neural Network (CNN), Model Optimization, Spatial Feature Mining