The traditional microblog advertisement filtering methods neglect the impact of factors such as data sparseness, semantic information, and advertisement background characteristics. Focusing on these issues, a new filtering method based on classification feature extension of Latent Dirichlet Allocation (LDA) was proposed. Firstly, microblogs were divided into normal microblog and advertising microblog, and the topic model of LDA was built respectively to infer the corresponding topic distribution, the words in the topic model were regarded as the basis of feature extension. Secondly, the background characteristics were extracted in conjunction with text category information during extension to reduce the impact on text classification. Finally, the extended feature vectors were served as the input of the classifier, and the advertisements were filtered depending on the results of Support Vector Machine (SVM) classification. In comparison experiments with the method only based on short text classification, the precision of the proposed method was averagely increased by 4 percentage points. The results indicate that the proposed method can effectively extend the text features and reduce the influence of background characteristics, it is more suitable for the filtering of microblog advertisement with great amount of data.