Skip to content

Schema — supervised (classification / regression)

Tables an experiment of type=supervised (or external) can publish. The profile discriminator is binary | multiclass | regression. All names below are suffixes of the canonical id; query the data-kb name (RULE 1), and confirm presence with the live catalog (RULE 6).

Contents

  • Fact: *__predictions
  • Dimension: *__rules (label rules, Train and Test measures)
  • Bridge: *__rules_fired
  • Dimension: *__archetypes
  • Example queries

Fact — *__predictions (PK row_index)

profile columns
common row_index (position), the input's declared key column(s) under their own names (catalog role key), row_role? (train / validation_holdout / outer_test / excluded — absent when the scored rows are not the fit's own dataset), archetype_id? (FK → archetypes)
binary predicted_proba (P of positive class), predicted_class (0/1), predicted_class_name
multiclass predicted_class (index), predicted_class_name, proba_<class> — dynamic wide columns, one per class; exact names only in the live catalog
regression predicted, predicted_lower_<level>? / predicted_upper_<level>? — one calibrated interval pair per configured level (predicted_lower_80, predicted_upper_95, …; exact coverage under a random split, approximate under grouped / temporal), residual? (actual − predicted, when y available)

Per-row errors and band coverage are honest only on held-out rows: filter row_role = 'outer_test'.

Dimension — *__rules (PK rule_id; classification only, when rules mined)

rule_id, engine, text (human-readable predicate), then the Train measures — on the rows the rules were found on, so optimistic: support (true positives / all rows — not the share of rows covered), precision, lift? (precision / base rate) — and the Test measures on held-out rows neither the model nor the rule search saw: support_holdout?, precision_holdout?, lift_holdout?, n_covered_holdout?, and holds_up? (holds_up / does_not_hold_up / too_few_rows; null without held-out rows), then class_key?. These are label rules — patterns in the labels found through the model's trees during MLI — the answer to "which rules apply / fire for predicted X?" Join via *__rules_fired (below). Quote the Test measures and prefer holds_up = 'holds_up'; call the Train measures in-sample. A separate, directly-applyable independent "Business Rules" model may also exist as an MLI artifact (not a SQL table) — see explainability (get_capabilities.has_business_rules).

Bridge — *__rules_fired (composite PK (row_index, rule_id))

row_index (FK → predictions), rule_id (FK → rules). Many-to-many: which rules fired on which rows.

Dimension — *__archetypes (PK archetype_id; present with MLI archetypes)

archetype_id, label, n_members, defining_features (short text), class_key?.

Standard joins

  • predictions ↔ input: the declared key column(s) (RULE 3).
  • predictions ↔ rules: predictions → rules_fired (row_index) → rules (rule_id).
  • predictions ↔ archetypes: archetype_id.

Example queries

Placeholders <input>, <predictions>, <rules>, <rules_fired> are the data-kb names for those tables.

Placeholder <key> is the input's declared key column (catalog role key).

Cohort profile of predicted-positive rows: 

SELECT s.region, COUNT(*) AS n, AVG(s.income) AS avg_income
FROM "<input>" s
JOIN "<predictions>" p USING (<key>)
WHERE p.predicted_class = 1
GROUP BY s.region ORDER BY n DESC;

Top rules by test lift among positives (rules that hold up on held-out rows): 

SELECT r.text, r.engine, COUNT(*) AS fired,
       r.lift_holdout, r.precision_holdout, r.lift AS lift_train
FROM "<predictions>" p
JOIN "<rules_fired>" b USING (row_index)
JOIN "<rules>" r      USING (rule_id)
WHERE p.predicted_class = 1 AND r.holds_up = 'holds_up'
GROUP BY r.text, r.engine, r.lift_holdout, r.precision_holdout, r.lift
ORDER BY r.lift_holdout DESC, fired DESC LIMIT 20;

Threshold what-if (no re-scoring): 

SELECT CASE WHEN predicted_proba >= 0.7 THEN 'flag' ELSE 'pass' END AS bucket,
       COUNT(*) AS n
FROM "<predictions>" GROUP BY 1;

Regression error by segment, on the held-out test rows: 

SELECT s.region, AVG(ABS(p.residual)) AS mae
FROM "<input>" s JOIN "<predictions>" p USING (<key>)
WHERE p.row_role = 'outer_test'
GROUP BY s.region ORDER BY mae DESC;

Regression interval coverage and width per level on the test rows (check each level's columns in the catalog first): 

SELECT AVG(CASE WHEN p.predicted + p.residual BETWEEN p.predicted_lower_80
                                               AND p.predicted_upper_80
                THEN 1 ELSE 0 END) AS coverage_80,
       AVG(p.predicted_upper_80 - p.predicted_lower_80) AS width_80
FROM "<predictions>" p WHERE p.row_role = 'outer_test';

Multiclass: read the actual proba_<class> column names from the catalog — do not hardcode proba_1.