The invention discloses a dynamic user attribute extraction method based on social media. The dynamic user attribute extraction method includes the steps: firstly, performing text preprocessing for an acquired training sample set and extracting subject terms to obtain K subjects and m subject terms of each subject; secondly, extracting short texts of a user to be processed, dividing time sub-segments, filling data through a time sliding window to obtain text data of the time sub-segments, counting the occurrence frequency of the subject terms after text preprocessing to obtain attribute weight information of the subjects, introducing time attenuation coefficient, sequentially acquiring user attribute features associated with time attributes in time sequence, extracting the user attribute features of the latest time sub-segment as current user attribute features and outputting the current user attribute features. Without external knowledge, the short texts of the social media are semantically expanded by disordered words in the texts, and dynamic user attributes can be extracted from micro-blog texts released or forwarded by users.