Classification and warehouse segmentation review project

Write two questions, not one vague model goal

The supervised task predicts a missed handoff at intake so dispatch can intervene. The unsupervised task asks whether warehouse profiles form stable operating groups worth investigating. They need different evidence. The classifier has labeled future outcomes, costs and a threshold. Clusters have distance, scale, membership stability and a proposed action test, but no automatic notion of accuracy. The logistic model and density clusters start these separate threads.

Freeze both information boundaries

Partition shipment records by site and time for classification, fit all transforms within training folds, and score an untouched later cohort once. For segmentation, define the reference sites and features before inspecting business outcomes used to judge actionability. A site’s future delay rate must not enter its historical cluster features. Record feature availability at the decision clock. The split guide and feature guide address the supervised boundary.

Compare models and decisions

Fit a training-rate baseline, logistic model, linear margin model and constrained tree. Select settings within development data, then compare precision, recall, log loss where probabilities exist, calibration and review workload on the same later cases. A margin score requires separate calibration before probability-based costing. The margin lesson makes that distinction. The final action threshold follows explicit false-alarm and missed-handoff costs.

Stress-test the segmentation

Compare scaled k-means, complete-linkage groups and DBSCAN over a declared range of reasonable parameters. Report unassigned DBSCAN sites, membership changes, cluster support and whether a segment label changes a staffing decision that can be tested prospectively. If the grouping collapses after a minor scaling change or lacks a specific action, keep it exploratory. Do not force every warehouse into a named segment merely to populate a dashboard.

Approve only a bounded pilot

The code checks whether the review packet has minimum evidence before a pilot discussion; its boolean fields need human-audited records behind them. Include model version, preprocessing state, inference p95, label maturity, owner, rollback and a monitored pilot period. A failed check holds the pilot without claiming the idea is worthless. Monitoring and release review give the follow-through.

Implementation

def pilot_readiness(packet):
    required = {
        "future_labels_sealed": "future evaluation",
        "training_only_transforms": "fitted transforms",
        "classifier_calibration_checked": "probability calibration",
        "threshold_costed": "decision cost",
        "cluster_stability_checked": "segment stability",
        "segment_action_defined": "segment action",
        "monitoring_and_rollback_ready": "operations",
    }
    missing = [label for key, label in required.items() if not packet.get(key)]
    return "pilot review" if not missing else "hold: " + ", ".join(missing)

warehouse_packet = {
    "future_labels_sealed": True,
    "training_only_transforms": True,
    "classifier_calibration_checked": False,
    "threshold_costed": True,
    "cluster_stability_checked": True,
    "segment_action_defined": False,
    "monitoring_and_rollback_ready": True,
}
assert pilot_readiness(warehouse_packet) == (
    "hold: probability calibration, segment action"
)

Performance and operating cost

For N shipment predictions and F linear features, one scoring pass is O(NF), while cluster comparisons can require O(S²) distance memory for S sites. The packet check is O(K) for K review requirements; it does not measure inference latency, clustering stability or policy value by itself.

Common Mistakes

Do not evaluate a classifier and an unlabeled clustering with one accuracy number.
Do not treat a margin score as a calibrated probability.
Do not give an unstable cluster a business label without an action test.