Protein interactions are of great biological interest because they orchestrate nearly all cellular processes and can further our understandings in biological processes and diseases. Protein interaction data like many real world datasets are imbalanced in nature. Most protein pairs belong to the non-interaction class and few belong to the interaction class. Most existing protein interaction prediction methods assume equal distribution of the positive and negative interaction data. In this study, we first analyze effects of various portions of negative samples on the performance of domain-based protein interaction prediction methods using Artificial Neural Network (ANN), Bayesian Network (BN), and SVM. Then we introduce cost-sensitive learning to address the class imbalance problem. Experimental results demonstrated that the addition of cost-sensitive learning to each classifier: ANN, BN, and SVM, indeed yields an increase in accuracy.