Machine Learning Overview (page 11 of 13)

With supervised problems, we want to balance bias and variance. Although not 100% correct, we can think of bias similar to accuracy (high bias = low accuracy and low bias = high accuracy). A high bias model makes assumptions that lead to underfitting. A low bias model makes assumptions that could lead to overfitting. Variance is how the predictions from the model differ between training and testing. The ideal model is low bias and low variance, but this is not realistic for many complex business scenarios. Reducing the bias often results in greater variance (vice versa).