classification method based on SVM is proposed for classification tasks of similar and incomplete multi-observation data. All single observation samples in a multi-observation set belong to a same class. In each classification, different assumptions about the class of multi-observation sets are made, two classification errors are obtained respectively for each assumption, and the final label is determined by comparing two classification errors. The proposed method takes advantage of the correlation within the same classes and the difference between different classes, and achieves the effective classification for similar and incomplete multi-observation data. Experimental results show the effectiveness of the proposed method.