vc_004 · Multi-table aggregation

Status: ✓ Passing · Last run: 2026-08-24T17:43:52 · Pipeline: v2 · History: 13/17 runs passed (76%)

What was tested

Two-table synthetic dataset: 500 clients + ~4,000 transactions joined on client_id. Target lives on the clients table. The signal must be recovered by aggregating transactions per client (count, mean amount, total). Verifies that the pipeline handles multi-row-per-entity sources via aggregate features rather than a row-multiplying equi-join.

What we planted in the data

• 500 entities
• Signal strength: 0.50 (sigmoid slope multiplier)

What the system did

• Training rows built: 500
• Features used in model: 10
• Model selected: LogisticRegression
• CV AUC: 0.750

Top features by importance:

• avg_txn_amount (0.561)
• total_txn_amount (0.270)
• max_txn_amount (0.173)
• segment_enterprise (0.040)
• signup_year_zscore_by_segment (0.036)

What we checked — all assertions passed

How this could have gone wrong

• If row_count_eq failed: Training df has exactly 500 rows (one per client). A naive multi-table join would inflate this to ~4,000 (one per transaction); the assertion catches that.
• If no_pipeline_errors failed: Pipeline runs cleanly with two-table input — no schema, join, or grain mismatches that would halt the run.
• If auc_in_band failed: AUC inside a band consistent with the planted signal: clients with higher mean transaction amounts are more likely to be high spenders, and aggregate features should recover that. Below 0.65 means the aggregations didn't fire; above 0.95 means something leaked the target column into features.

Why this case matters

Multi-table is the failure shape the panel-fan-out fix had to handle generically: when a source table has many rows per entity, the pipeline must aggregate it (correlated-subquery features with time bounds), not blindly equi-join it into pred_grid (which would inflate training rows and leak future transactions into the features for any given snapshot). This case anchors that behavior on a clean two-table synthetic.

Reproducing

# from auto_insight_api/
python -m validation.v2 run vc_004 --pipeline v2 -v

• Case config: validation/v2/cases/04_multi_table.yaml
• Data shape: multi_table
• Analytics type: ml_binary
• Related: feature_materialize: 'event-stream source skipped (multiple rows per entity)', fan-out hard gate at pred_grid_plus

← back to validation report