Protein binding hot spots are those residues that locate at the interfaces of protein-protein interaction, which can influence the interaction of proteins significantly, although they consist of a small part of the interface residues only. Since traditional experimental methods for prediction of hot spots are quite complex and time-consuming, then recently, bioinformatics methods are adopted as efficient tools in the area of hot spots prediction at protein-protein interface. In this paper, a novel prediction model is proposed for the prediction of hot spots at the protein-protein interaction interfaces based on the Extreme Learning Machine (ELM) and a new way for feature selection. Different from existing methods, the selection of the classifier in our method included two parts: the first part aimed for deleting some redundant features from the original features without prediction, and the second part aimed for constructing our final prediction model based on the prediction. Our major contribution was that the ELM and a new prediction model were introduced into the area of hot spots prediction, which could improve the prediction performance remarkably, when comparing with some traditional existing methods. And the simulation results based on two benchmark datasets ASEdb and BID showed that the newly proposed prediction model outperformed some of the existing well known methods such as the Robetta, KFC, and HotPoint, etc.