Peptide-binding proteins are excessive in living cells and proteinpeptide interactions mediate a wide range of cellular functions. Prediction of protein-peptide binding residues has been vital and popular in the past decades and machine learning methods have gained more attention in recent years. However, the data imbalance problem has not been dealt with effectively. On this matter, we study the effects of sampling methods and degrees of imbalance on data classes on construction of prediction model. We first developed the NearMiss under-sampling method (NMUS) as a way to screen out a given number of quality data samples from majority class to balance the data sets. The remarkable sensitivity (SEN) with 0.818 shows the advantage of NMUS in handling class imbalance problem. This research carried on valuable analysis on data imbalance problem and achieved a better prediction of protein-peptide binding interaction.