FIGURE 2.

The graph depicts the different workflows in traditional statistics compared with machine learning approaches. In traditional statistics, one approaches a dataset with predefined assumptions, reduces the entire dataset according to those assumptions and then tests a certain hypothesis for significance. In contrast, unbiased machine learning approaches split the dataset into training and test data and let an algorithm learn from the training data in an unbiased and hypothesis-free manner. The evaluation of the analysis then depends on how well the model performs when applied to the test data. These two approaches can yield quite different results.