Real estate prices are affected by locational heterogeneity, locational autocorrelation and nonlinearity of locational distribution, which makes it challenging to evaluate housing prices quickly, massively and accurately under complex urban locations in China. Based on trend surface analysis and Bayesian optimization modeling, this paper constructs the BO-TSA-XGBoost model, which converts the valuation problem into an attribute space partitioning problem, to achieve the identification and learning of locational information of real estate mass appraisal in a complex data environment. In this paper, 34 460 second-hand house transaction data from all 16 administrative districts of Shanghai from 2020 to 2021 are collected for empirical analysis. The research results show that the BO-TSA-XGBoost model can accurately identify and learn location information under complex data, effectively solve the problem of evaluation accuracy degradation caused by complex locality characteristics, and achieve a high level of accuracy in multiple price intervals and multiple complex locations; The trend surface analysis and Bayesian optimization algorithm significantly improve the evaluation accuracy and evaluation robustness of the original ensemble learning model; housing location information is the key to the mass appraisal model.