Rapid Recall Key 41–60
- Recombination risk: combining sources or narrow outputs can recreate identifying/sensitive information not visible in inputs alone.
- Before integration: inspect/profile; capture Metadata; assess DQ/reliability; understand semantics and relationships.
- Why DQ early: it determines source usability and mapping/alignment; doing it last embeds poor evidence and obscures root causes.
- Master/Reference Data: provide trusted identifiers/domains/context to align observations consistently across sources.
- Model vs input quality: sound hypothesis can fail with bad evidence; excellent data cannot rescue invalid assumptions—evaluate both.
- Why historical pre-population: history supplies examples/baselines for model fitting and comparison rather than beginning with an empty current stream.
- Training: repeatedly run/fit the model against data, adjust it, verify assumptions, and improve fit/performance.
- Over-fitting: model learns training-specific noise/patterns too closely and fails on unseen data.
- Train vs validation: training fits; validation chooses/tunes alternatives and estimates selection performance.
- Validation vs test: validation influences choices; test remains independent for final generalization evidence.
- Test reuse: when results influence model/feature choices, the test set joins selection and its error estimate becomes optimistically biased.
- K-fold: rotate held-out subsets/folds so each can serve as validation, improving robustness with limited data.
- Outliers: may be valid, rare, high-risk, or the target signal; rarity alone is not error.
- Fit visualization: answer a defined question for a defined audience clearly, neutrally, with context/assumptions needed for correct interpretation.
- Static vs interactive: fixed view vs consumer-explorable/manipulable view.
- Deploy when: useful, business-relevant, statistically/ethically/technically sound enough, feasible to operate and monitor.
- Monitor: model performance, value, source changes, DQ, operational health, latency/cost, adoption, and new requirements/questions.
- MPP: parallel architecture partitioning data/computation across nodes with dedicated resources for scalable analytical processing.
- Distributed file-based: distributed environment flexibly storing/landing very large structured, semi-structured, and unstructured data across nodes, often at lower cost.
- MapReduce: Map → Shuffle → Reduce.
← Prompts 41–60 · Reconstructions →