Using appropriately set hyper-parameters, many machine learning algorithms can optimally learn how to map your data with respect to your target outcome. Yet, their predictive performance can be improved further by fixing hidden and subtle problems in data. It is not simply a matter of detecting any missing or outlying case. Sometimes, it is a matter of whether there are any groups or unusual distributions in the data (for instance, multimodal distributions). Clearly drafted data plots can explicate the relationship between variables, and they can lead to the creation of new and better features in order to predict, with increased accuracy, your target variable.
The just-described practice is called ...