Binning and Discretization

Binning takes a continuous variable and groups its values into a smaller number of discrete intervals, or bins. The three common approaches differ in how the bin boundaries are chosen. Equal-width binning divides the variable's range into intervals of the same size, such as every 10 years for age, which is simple but can leave some bins nearly empty and others overcrowded if the underlying data isn't uniformly distributed. Quantile binning, also called equal-frequency binning, instead chooses boundaries so that roughly the same number of data points fall into each bin, which handles skewed distributions better, though the boundaries become data-dependent and can shift if the dataset changes. Supervised binning uses the target label to choose boundaries that best separate outcomes, for example finding income cutoffs that most cleanly split loan defaulters from non-defaulters, often using decision-tree-like splitting criteria. Binning can help models in specific ways: it can reduce the impact of measurement noise and outliers by treating a range of values identically, it can let linear models capture non-linear relationships between a variable and the outcome, since each bin gets its own coefficient, and it produces easier-to-interpret rules for humans, like risk tiers or age brackets. However, binning always discards information, since values within a bin become indistinguishable, and poorly chosen boundaries, particularly narrow bins near a decision boundary, can reduce accuracy or introduce artificial discontinuities in what was originally a smooth relationship. A common misconception is that binning is always needed for tree-based models; modern tree algorithms like gradient boosting already handle continuous splits internally and often don't benefit from pre-binning, though some implementations use internal histogram binning purely as a computational speed optimization.