Prediction of hot spots in protein interfaces provides crucial information for the research on protein-protein interaction and drug design. Existing machine learning methods generally judge whether a given residue is likely to be a hot spot by extracting features only from the target residue. However, hot spots usually form a small cluster of residues which are tightly packed together at the center of protein interface. With this in mind, we present a novel method to extract hybrid features which incorporate a wide range of information of the target residue and its spatially neighboring residues, i.e. the nearest contact residue in the other face (mirror-contact residue) and the nearest contact residue in the same face (intra-contact residue). We provide a novel random forest (RF) model to effectively integrate these hybrid features for predicting hot spots in protein interfaces. Our method can achieve accuracy (ACC) of 82.4% and Matthew's correlation coefficient (MCC) of 0.482 in Alanine Scanning Energetics Database, and ACC of 77.6% and MCC of 0.429 in Binding Interface Database. In a comparison study, performance of our RF model exceeds other existing methods, such as Robetta, FOLDEF, KFC, KFC2, MINERVA and HotPoint. Of our hybrid features, three physicochemical features of target residues (mass, polarizability and isoelectric point), the relative side-chain accessible surface area and the average depth index of mirror-contact residues are found to be the main discriminative features in hot spots prediction. We also confirm that hot spots tend to form large contact surface areas between two interacting proteins. Source data and code are available at: http://www.aporc.org/doc/wiki/HotSpot.