Often a dataset embodies some form of bias either due to the previous human decsions embodied in them, or due to the way in which they were collected. Techniques to de-bias attemto to undo these effects on the data. For example, the weighting assigned to items during training may be adjusted to account for poorly represented types of people. It is important to realise that these de-biasing techniques are almost always partial, and may introduce their own problems including introducing different bises.
Used on page 487