Online media is an important source for sentiments exposed by individuals on goods, services, organizations, and other objects of interest. While firms can benefit from using these sentiments for decisionmaking, the classification of sentiments is difficult because of volume, velocity, and variety. Machine learning is an effective technique for sentiment classification, which neither requires formalized knowledge about the domain nor the language used. Although the literature provides a rich body of classification methods, system designers and researchers still face the problem of reasonably selecting designs. In this paper, we seek to contribute to the understanding of machine learning for sentiment classification. We report an experimental study that tests the effects of three design factors, i.e., text representation, feature weighting, and machine learning algorithm, on accuracy. The findings can be useful for empirically informed classifier design.