Log and Box-Cox Transformations

A transformation changes the numerical scale of a variable to make patterns easier for a model to represent. A logarithm grows slowly for large positive values, so it can reduce right skew and turn multiplicative relationships into additive ones. The change also means differences on the transformed scale no longer correspond to equal absolute differences on the original scale. The ordinary logarithm requires positive inputs. A log1p transform uses log(1 plus x) and accommodates zero but still requires x greater than negative one. Adding an arbitrary constant to permit negatives changes the interpretation and should be justified. Box-Cox considers a family of power transformations and estimates a parameter, but its standard form requires strictly positive values. Yeo-Johnson is another power-transform family that can handle zero and negative values. These methods can help with skewed features or targets, but they are not universal preprocessing requirements. Tree-based models may be less sensitive to monotonic rescaling than models based on linear relationships or distance, though distribution changes can still affect some workflows. A transform does not remove outliers, establish normality, fix biased sampling, or guarantee better predictions. Compare model behavior and errors on held-out data. For target transformations, fit the transformation parameters on training data and transform training targets accordingly. At prediction time, apply the inverse transformation to return to the original units. Because nonlinear inverses do not preserve averages, simply exponentiating an average predicted log value can produce a biased estimate of the original-scale conditional mean. The appropriate correction depends on the modeling assumptions and evaluation objective. Use a pipeline to avoid fitting transformation parameters using validation or test examples. Record the transform and fitted parameter so inference uses the same mapping. Report metrics in a scale that matches the practical task, and inspect residuals and subgroup effects rather than judging only the transformed distribution.