It is very important to reduce the candidate features in the machine learning such as classification and clustering. Most of the existing methods are based on a single feature on the target T or the association between the feature and the feature on the Y. However, these methods do not take into the combined features, such as attributes A, B contains a little amount of information in Y, and even completely independent of Y, but A & B can provide information on Y lot of information, or even completely determine the Y. Based on this, we can extract an algorithm to find single and combined features from the feature set, firstly combination of non-significant features in accordance with the conditional probability distribution table to generate new candidate features Then, the single feature and the combined features are chosen based on the criterion of the maximum correlation and the minimum redundancy. Finally, the experiment is carried out on the virtual and real data sets respectively, and the experimental results show that the feature selection algorithm can mine the dataset better, Which improves the accuracy of the corresponding machine learning algorithm to a certain extent.