In order to help the pig farm managers better carry out the reproductive management of sows, predict the high and low yield sows, and timely eliminate the low yield sows, in this study, we collected and sorted out the dataset of three sow populations, including birth herd, farrow herd, breed and birth weight of different parities, formulated the classification standard of sow high and low yield, and used boruta package in R software to screen out the important characteristics affecting high and low yield of sows. Four different machine learning methods, logistic regression (LOG), decision tree (DT), random forest were (RF) and support vector machine (SVM) were used to construct the classification model of high and low yield sows, and the decision tree view analysis was carried out to explore the related factors affecting the highest yield of sows. The results showed that the classification accuracy of the four machine learning methods for sow high yield classification model was about 71%, and the highest was 84%. It was also found that SVM as the best modeling method appears most frequently across all data sets and different classification criteria, followed by LOG and DT. The decision tree view showed that birth herd, breed and birth weight of different parties were important leaf nodes for dividing the highest yield sows, and these characteristics can be used to predict the most productive sows, with an accuracy of 73%-82%. These results indicated that it will be a good choice to use machine learning method to predict the high and low yields of sows in the future.