Classification Drill — 20 Fast Decisions
Classify before looking at the key.
- A model estimates next-month default probability.
- The system recommends the credit-limit action expected to reduce loss.
- Billions of rows overwhelm current storage/processing.
- Sensor events must be acted on within seconds.
- Two files use different identifiers and structures and are hard to combine.
- A provider's data cannot be trusted.
- Labeled fraud/not-fraud examples train a classifier.
- No labels; the team wants natural customer groups.
- An agent improves choices using reward feedback.
- Find products frequently purchased together.
- Group similar customers without predefined segments.
- A raw-data repository has no source/provenance/inventory and users cannot find data.
- A source is well documented but 40% of a critical key is missing.
- A source is complete but nobody knows origin or lineage.
- A model is tuned every time the “test” score changes.
- Training accuracy is high but unseen-data performance is poor.
- Rare transactions are the possible fraud cases being sought.
- A plant can act only once per day, but a vendor proposes millisecond analytics.
- Two anonymous sources together identify small groups.
- Petabytes of mixed raw logs need cheap flexible landing, not immediate complex SQL.
Key
- Predictive analytics
- Prescriptive analytics
- Volume
- Velocity; likely speed/operational-analytics requirement
- Viscosity (Variety supports if formats differ)
- Veracity
- Supervised learning
- Unsupervised learning / clustering
- Reinforcement learning
- Association / data mining
- Clustering / unsupervised mining
- Data-swamp risk / Metadata failure
- Data Quality problem
- Metadata problem
- Test-set independence failure; use validation for tuning
- Over-fitting/generalization failure
- Do not auto-delete; investigate as possible target signal
- Latency/cost mismatch; business does not justify speed architecture
- Recombination privacy risk
- Distributed file-based environment