Feature Selection#
Feature profiling, lineage and drift#
scikitplot.feature_selection helps with tables that have too many columns
to reason about one by one. It answers three questions that are usually
answered by hand and forgotten:
What is each feature? Role, dtype family, semantic kind, user tags, a bin recommendation, and policy-backed assessments such as skewed, discrete, constant, binary or identifier-like, each with evidence and an explicit
unknown/not_applicablestate.What happened to each feature in a fitted pipeline? Which steps touched it, which columns it became, where it was dropped, and how a model’s importances map back to raw columns.
Has it drifted? A frozen training reference compared with later data.
Important
Profile the training split only. Validation, test and production data are audited against the fitted schema and the frozen reference; they are never profiled into them.
Profile a table, ask features about themselves, group and select them.
Policies in TOML, YAML or JSON; thresholds; overrides; custom assessments.
Route columns in a ColumnTransformer, trace a fitted pipeline, attribute importances.
Freeze a training reference and judge later data feature by feature.
Examples#
See Feature Selection for executable examples that accompany this guide.