Value grouping

What it represents

Real CRM picklists drift. One industry ends up stored as “Software”, “software”, “SOFTWARE ” with a trailing space, and “Software / Services”. To a person those are one value. To a model they are four, each holding a quarter of the evidence, each too thin to be worth much.

Value grouping collapses spelling variants of the same value into one group, one field at a time, and keeps a map of what merged into what.

How it’s calculated

The grouping pass is deterministic. For every distinct value in the field it applies Unicode normalization, folds case, trims and collapses internal whitespace, treats ampersands, plus signs, and slashes as separators and collapses runs of them, and strips separators from the ends. Values that come out identical merge into one group. The group’s label is the spelling that appears most often in your own data, so what you see is always a real value from your CRM.

Very rare values are handled separately: below a frequency floor they collapse into a single tail level in the model’s feature space, without changing the map you look at.

Grouping is per field and opt-in. A map is created as a draft and changes nothing until you activate it, and each activation gets its own version. That version is recorded with any model trained afterwards, so a model always scores with the exact map it learned on. Activating a new map does not retroactively alter an existing model’s behavior.

What it deliberately does not do: it does not merge values that only a reader would recognize as the same. “Web Seminar” and “Webinar - Q3” survive as separate groups, because a rule that guessed at meaning would be a rule you could not audit.

What it means for you

Worked example: a lead-source field holds 51 distinct values. The deterministic pass merges casing, spacing, and separator variants down to 34 groups, and the long tail of one-off values collapses into a single level. The model now sees denser levels with more evidence behind each. The genuinely semantic pairs still sit there as separate groups. ax1om does not merge on meaning today, so when two of those spellings really are one value, the cleanup belongs in your CRM.