A large number of protein-protein interactions (PPIs) have buried in massive biomedical articles published over the years. This leads to the development of automatic PPI extraction methods. However, existing methods based on supervised machine learning still face some challenges: (1) the feature space exploited in these methods is very sparse; and (2) the data used for training are imbalanced with respect to categories to be classified. In this paper, we first construct rich and compact features to alleviate the issue of feature sparseness. With these features, our method outperforms baselines by up to an F-score of 9.58% on the original AIMed corpus. Furthermore, we propose a data sampling strategy based on under-sampling to address the class imbalance problem. In order to re-balance data distribution, samples of the majority class are removed according to the prediction results iteratively. By this means, our method achieves a further 2.49% improvement in F-score on the original AIMed corpus.