An Evolutionary Hyperparameter-Optimized Off-Policy Feature Selection for Real Estate Tax Prediction

Abstract Precise predictions of real estate taxes (RETs) are critical for property owners and government agencies because they influence economic planning and public revenue generation. Traditional deep learning methods for RET often struggle with feature selection and are highly sensitive to hyperparameter settings. To overcome these issues, we present a sophisticated predictive model for RET that incorporates a refined reinforcement learning (RL) technique, termed off-policy proximal policy optimization (PPO), which excels in feature selection, and with an advanced differential evolution (DE) algorithm for optimal hyperparameter tuning. The off-policy PPO component continually updates its feature selection based on data interactions, sharpening its focus on crucial features and preventing overfitting. Moreover, the DE algorithm has been enhanced with a novel mutation strategy that uses k-means clustering to identify key clusters. We tested our model on diverse real estate datasets from Kaggle, covering locations such as Bucharest, California, Helsinki, King County, and Saudi Arabia. Our findings show that our model significantly outperforms traditional pricing tools, with mean absolute percentage error (MAPE) ranging from 0.202 to 1.362 across various datasets. This superior performance underscores the broad applicability of the model across markets, markedly improving financial analysis and strategic planning.