Protein mass spectra pattern recognition is a new forum in which many machine learning algorithms have been conducted to enhance the chance of early cancer diagnosis. The high-dimensionality-small-sample (HDSS) problem of cancer proteomic datasets still requires more sophisticated approaches to improve the classification accuracy. In this study we present a simple ensemble strategy based on measuring the generalizing capability of different subsets of training data and apply it in making final decision. Using a limited number of biomarkers along with 5 classification algorithms, the proposed method achieved a promising performance over a well-known prostate cancer mass spectroscopy dataset.