Data transformation changes values into a form suited to analysis. For a pandas Series of categorical labels, `map()` applies a mapping to each value and returns the transformed values. Assign the result back to the column when the numeric codes should replace the original labels. section 5.2.1
Encode a binary outcome
To prepare raisin varieties for logistic regression, define `class_mapping = {'Kecimen': 0, 'Besni': 1}` and run `raisins['Class'] = raisins['Class'].map(class_mapping)`. Check selected columns with `head()` and `tail()`: the former Kecimen rows should show 0, and the former Besni rows should show 1. Either assignment of 0 and 1 works, but interpret predictions as the probability of the class coded 1. section 5.2.1
The mapping must serve the meaning of the variable, not just its data type. In fourth-down play data, categorical venue and play type become numeric inputs for decision trees and random forests. Score differential needs a different transformation: multiply it by −1 for road-team plays so its sign describes the advantage from the offense’s perspective. section 6.2.2
>>> OFFENSIVE_TEAM_VENUE_mapping = {'Road': 0, 'Home': 1}
>>> nfl_pbp['OFFENSIVE_TEAM_VENUE'] = \
>>> nfl_pbp['OFFENSIVE_TEAM_VENUE'].map(OFFENSIVE_TEAM_VENUE_mapping)
>>> PLAY_TYPE_mapping = {'Run': 0, 'Pass': 1}
>>> nfl_pbp['PLAY_TYPE'] = \
>>> nfl_pbp['PLAY_TYPE'].map(PLAY_TYPE_mapping)
The listing shows two mapping dictionaries and the corresponding Series `map()` assignments for venue and play type. section 6.2.2
Verify transformed columns before modeling. The football data’s `info()` output reports 881 non-null venue values among 883 rows after conversion, so numeric dtype alone does not establish that every row has a usable value. Similarly, replacing missing yards gained with zero is appropriate only when zero reflects the play; imputation can otherwise distort predictions. section 6.2.2
Numeric transformations address another problem: distribution shape. Logarithmic, reciprocal, and square-root transformations can compress right-skewed values before linear regression, depending on the distribution. A fitted relationship then applies on the transformed scale, not automatically on the original scale; consider whether a different model better represents a nonlinear relationship. section 4.1.3
