Currently, Android malware detection is one of the hotpots in the security research field. Since Android is open source and very popular, the Android platform becomes a target of most malwares. Current approaches only extract syntax features or semantic features respectively so that it is difficult for them to know the real intention of the malware exactly. We propose a hybrid feature extraction method, using the set of class-based taint propagation paths as semantic features and claiming permissions and Intent-Actions as syntax features. We normalize all the extracted features before training and clustering data sets by K-means, and then produce feature vectors of each malware family. Finally we adopt the Euclidean distance computation to measure the similarity between the unknown program and feature vectors. The prototype is implemented on top of FlowDroid to analyze 400 real programs, and the results demonstrate that the method has higher precision.