One-Hot Encoding

Column Order Is Part of the Data

One-hot columns are only meaningful when the column order is kept fixed.

highlighted = computed this step

Column order is part of the encoding

The same indicators only make sense with the pinned order red, blue, green, other.

(red,blue,green,other)(\text{red},\text{blue},\text{green},\text{other})
Categorical Features ExactlyExact one-hot columns from a pinned category list.one-hot columns from pinned category orderrowcolorredbluegreenotherAred1000Bblue0100Cgreen0010category order: red, blue, green; other reservedpreprocessing rewrites category tokens into indicator columnsunknown category uses the pinned other-column policyNOT training; NOT learning; NOT generalizationNOT accuracy; NOT calibration; NOT a guarantee of model qualityNOT probability truth

Do not silently reorder later

If a later page reads the second column, it must still mean blue. The order is a visible part of the preprocessing rule.

second column=blue\text{second column}=\text{blue}
Categorical Features ExactlyExact one-hot columns from a pinned category list.one-hot columns from pinned category orderrowcolorredbluegreenotherAred1000Bblue0100Cgreen0010category order: red, blue, green; other reservedpreprocessing rewrites category tokens into indicator columnsunknown category uses the pinned other-column policyNOT training; NOT learning; NOT generalizationNOT accuracy; NOT calibration; NOT a guarantee of model qualityNOT probability truth

The renderer checks the order

The diagram is fail-closed: changing the category order or the column order is rejected before rendering.

order pinned at the trust boundary\text{order pinned at the trust boundary}
Categorical Features ExactlyExact one-hot columns from a pinned category list.one-hot columns from pinned category orderrowcolorredbluegreenotherAred1000Bblue0100Cgreen0010category order: red, blue, green; other reservedpreprocessing rewrites category tokens into indicator columnsunknown category uses the pinned other-column policyNOT training; NOT learning; NOT generalizationNOT accuracy; NOT calibration; NOT a guarantee of model qualityNOT probability truth