When considering potemtial bias, a proxy indicator is some combination of features that reveal the value of a protected characteristic such as gender or ethnic background. For example, Days taken as holiday from work might be indicative of a person's religion, or subjects tkane at school may be related to gender. Often it is assumed that if protected characteristics are removed form a dataset before it is used for training, then the reuslying machine learning system will be free of bias. However, proxy indicators mean that decisions are still effectievly made based on the protected characteristic.
Used on page 491