Protein-protein interactions (PPIs) are the basis to interpret biological mechanisms of life activity, and play vital roles in the execution of various cellular processes. The development of computer technology provides a new way for the effective prediction of PPIs and greatly arouses people's interest. The challenge of this task is that PPIs data is typically represented in high-dimensional and is likely to contain noise, which will greatly affect the performance of the classifier. In this paper, we propose a novel feature weighted rotation forest algorithm (FWRF) to solve this problem. We calculate the weight of the feature by the chi(2) statistical method and remove the low weight value features according to the selection rate. With this FWRF algorithm, the proposed method can eliminate the interference of useless information and make full use of the useful features to predict the interactions among proteins. In cross-validation experiment, our method obtained excellent prediction performance with the average accuracy, precision, sensitivity, MCC and AUC of 91.91, 92.51, 91.22, 83.84 and 91.60% on the H. pylori data set. We compared our method with other existing methods and the well-known classifiers, such as SVM and original rotation forest on the H. pylori data set. In addition, in order to demonstrate the ability of the FWRF algorithm, we also verified on the Yeast data set. The experimental results show that our method is more effective and robust in predicting PPIs. As a web server, the source code, H. pylori data sets and Yeast data sets used in this article are freely available at http://202.119.201.126:8888/FWRF/.