As the carrier of human activity and social development, buildings are the most important geographical entities that constitute the spatial structure of a city. It is one of the urgent tasks in the construction of smart cities in China to build elaborate digital models of urban buildings. Classifying a large amount of buildings by their functions facilitates urban functional area division and urban spatial cognition, thus assisting the government in population estimation, land management, urban planning, and smart city construction. In this paper, POI (Point of Interest) with rich semantic information including name, address, and types was used as the main data source, because it was more accessible and updated more frequently than the traditional geographic information data. The process of finding out the functional type of a building was similar with identifying urban functional areas by using POI data, but there existed the problem of low classification rate due to the sparsity of POI. Therefore, to improve the traditional quantitative identification of urban functional areas, this study attempted to calculate the weighted frequency density ratio of each type of POIs inside and within a certain range around a building. Experimenting on more than 5000 buildings near South Shawo Bridge in the west of Beijing, the study found that 93.04 percent of the buildings were effectively classified into different functional types: residential, commercial, public service, and other three mixed types. The classification rate has been greatly improved compared with that of the traditional method. These classified buildings showed the spatial distribution of functional areas more clearly and precisely than blocks used in identifying urban functional areas, since too many multi-functional blocks with very limited practical meaning were identified by using the traditional method. In order to calculate the classification accuracy, more than 2000 randomly selected buildings were manually divided into functional classes with the assistance of POI and AOI data. The overall classification accuracy reached 91.18 percent compared with the manually classified result. The classification error was mainly caused by the shortage of POI and the poor data quality, which could be avoided by merging multi-source POI to improve the data quality or applying various Internet location information, such as the social media data and the real estate transaction data. However, by using easily accessible web POI data, the proposed method, which can replace manual classification in an automated way, has greatly improved the effectiveness of classifying large number of buildings into different functional types, and shown higher accuracy than existing researches.