Machine learning is no silver bullet. You still have to understand the algorithms and the assumptions they make. For example, if you're using a support vector machine, do you have reasonable to believe your data is linearly separable? If you're using a kernel, does the kernel make sense for your data? You can't just blindly try methods till something seems to work, because you'll end with something that works by accident on your test data but doesn't apply to the real problem you're trying to solve. Understand the data and everything else follows.