Machine Learning techniques are widely used in many areas including mobile applications. Currently, these techniques are often applied as a black box. That is, most researchers in the mobile area simply get one machine learning package, conveniently run it on their data, and report the result. However, without getting into certain details of the machine learning method, not only is the obtained performance far from the optimum, but also the reported results may be misleading. In this paper, through a case study on porting an SVM-based transportation-mode detector to a low-power and low-memory device, we demonstrate the importance of being careful in applying machine learning methods for mobile applications.