PRODUCTION SCENARIO
A distribution center has months of normal package-scan records but no reliable labels for damaged parcels. It wants to route unusual scans to a human inspector.
Which ML approach will meet these requirements?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Use an anomaly detection model to score how unusual each new scan is.
Anomaly detection learns what normal looks like from unlabeled records and scores each new scan against it, which is what routing unusual scans to an inspector needs. Fine-tuning a model to classify damage still needs labeled damaged and undamaged examples, which the center does not have. Fixed thresholds work one field at a time and miss a scan that is unusual only in the combination of its values. Clustering produces no per-scan score, and the smallest cluster may be a legitimate package type.
Official Source: https://docs.aws.amazon.com/sagemaker/latest/dg/algos.html ↗