Buildings account for a large amount of energy consumption in urban areas. Understanding the impacts of urban neighborhood characteristics on building energy consumption can help identify ways of how to improve neighborhood energy and environmental performance towards a more sustainable built environment. Despite the availability of many building energy consumption assessment/prediction models and tools, the majority of existing efforts mainly focus on the building scale; and the small number of neighborhood-scale analyses are typically conducted using building-scale simulations instead of taking a data-driven approach. To address these gaps, this paper proposes a hybrid machine-learning approach (Clus-SVR) that combines a support vector regression (SVR) algorithm with a clustering algorithm for predicting the building energy consumption at the building scale and the neighborhood scale based on the socioeconomic and building characteristics of the neighborhood (e.g., per capita income, average gross floor area, etc.). As a preliminary study, the proposed Clus-SVR algorithm was tested in predicting building-scale and neighborhood-scale consumption using the energy consumption data of the City of Chicago, and was evaluated in terms of mean absolute percentage of error (MAPE) and coefficient of variation (CV). The paper discusses the proposed algorithm and the performance results, and identifies the impacts of urban neighborhood features on building energy consumption levels.