Hershel Safer

Weight of evidence (WOE) is used to assess the relative risk of the values taken by a categorical feature. For example, suppose that the goal is to predict the probability of default on loan. One feature might be homeownership, with values "own outright," "own with a mortgage," "rent," "live with parents," and "other."

With unordered, non-numeric values, a feature like this cannot be used directly with most machine-learning methods. WOE addresses this challenge by converting each value, say "own with a mortgage," to the relative risk of an applicant in the group with that value paying off the loan. Specifically, WOE is the log odds of the categorical value being associated with a negative target value.

When the WOE of a categorical value is positive, the probability of the loan being paid in full is above average among all applicants, and vice versa when WOE is negative.

WOE has advantages over the one-hot encoding that is commonly used for categorical features. WOE avoids the proliferation of features when one-hot encoding is used with categorical features that take many values. WOE also does not lose information the way that one-hot encoding does when tree-based algorithms select a subset of features at each node.