Drift
A major determinant of data quality is drift. Drift (also: dataset shift) means that the patterns in data change over time. Drift is important because the performance of a machine learning model can be adversely affected by changes to the dataset.
Drift transitions can occur abruptly, incrementally, gradually, or be recurring. This is illustrated here:
Figure 8.4: Four types of concept drift transitions
When the transition is abrupt, it happens from one time step to another without apparent preparation or warning. In contrast, it can also be incremental in the sense that there's first a little shift, then a bigger shift, then a bigger shift again.
When a transition happens gradually, it can look like a back and forth between different forces until a new baseline is established. Yet another type of transition is recurring or cyclical when there's a regular or recurring shift between different baselines.
There are different kinds of drift:
...
