Bias occurs when a model produces consistently unfair or inaccurate results. This can arise in two contexts:

  • Algorithmic/statistical bias: The error due to overly strong or simplistic assumptions about the data. High bias leads to underfitting, where the model cannot capture the true relationship between inputs and outputs.
  • Data/ethical bias: The unfairness or systematic skew in predictions caused by biased data collection, feature choices, or training design.

In the bias–variance framework:

  • Bias = the error from the model being unable to learn the true mapping between inputs and outputs.

Related: