The invention discloses a website classification method based on a network relation graph. The website classification method comprises the following steps: using the malicious data of 'security alliance' as a sample data source and obtaining existing data classification and URL data; forming a site mapping table and a word frequency analysis table, extracting a fingerprint feature of a sample for constructing a feature table and forming a type table; forming an undirected weight graph by each site and a connection weight thereof; constituting a network relation graph by numerous sites and weights, and dividing the big network relation graph into several sub-graphs through a map clustering algorithm; extracting and classifying the fingerprint features of each task unit on a server thereof through a site fingerprint feature extractor and a site fingerprint feature classifier. According to the website classification method based on the network relation graph, the detection speed of website classification is remarkably improved, and the specific type is distinguished. For the newly added site, a processing unit is determined according to the connection relation of the newly added site and the existing site, the computational complexity and the number of iterations of the training are thus effectively reduced, and the computing resources are saved.