Currently, the number of mobile malware programs is explosively growing, and the increasingly large feature library poses challenges to security solution providers. Traditional detection methods cannot deal with the huge amount of data promptly and effectively. Mobile malware detection methods based on machine learning have problems of excessive numbers of features, low detection accuracy and unbalanced data. In this paper, a feature selection method based on the mean and variance of samples was proposed to reduce the features without affecting classification. Different feature extraction algorithms were implemented to construct an ensemble learning model for high detection accuracy, including Principal Component Analysis, Kaehunen-Loeve Transformation and Independent Component Analysis. To solve the problem of unbalanced data of Android App samples, a multi-level classification model based on the decision tree was also developed. Experimental results show that the proposed methods can detect Android malware effectively, and the accuracy is increased by 6.41%, 3.96% and 3.36%, respectively.