Protein-protein interaction (PPI) plays crucial roles in the performance of various biological processes. A variety of methods are dedicated to identify whether proteins have interaction residues, but it is often more crucial to recognize each amino acid. In practical applications, the stability of a prediction model is as important as its accuracy. However, random sampling, which is widely used in previous prediction models, often brings large difference between each training model. In this paper, a Predictor of protein-protein interaction sites based on Extremely-randomized Trees (PETs) is proposed to improve the prediction accuracy while maintaining the prediction stability. In PETs, a cluster-based sampling strategy is proposed to ensure the model stability: first, the training dataset is divided into subsets using specific features; second, the subsets are clustered using K-means; and finally the samples are selected from each cluster. Using the proposed sampling strategy, samples which have different types of significant features could be selected independently from different clusters. The evaluation shows that PETs is able to achieve better accuracy while maintaining a good stability. The source code and toolkit are available at https://github.com/BinXia/PETs.