Abstract: With the acceleration of global urbanization, the importance of seismic hazard assessment for urban building clusters has become increasingly prominent. When the number of buildings is very large, traditional methods of seismic damage statistics and simulation struggle to strike a balance between computational efficiency and precision. The ongoing advancement of machine learning technologies offers a novel approach to address this paradox. Among these, Collaborative Filtering (CF), a well-established recommendation system algorithm, has demonstrated success in handling massive data sets. Leveraging analogy, this study proposes a CF-based framework for predicting the nonlinear seismic response of building clusters, enabling efficient computation of large-scale nonlinear responses under batch seismic actions. Through an analysis task involving 10,000 buildings within a specific urban area subjected to 3052 seismic actions, the paper delineates the process of establishing the CF prediction model. The results indicate that the model, trained with a minimal quantity of computational results as samples, achieves prediction accuracy within a Mean Absolute Percentage Error (MAPE) of less than 5%. Compared with conventional dynamic time-history analysis methods, the prediction time required by the model is significantly reduced by several thousand hours.