Deep Battle Cards G–L
G — Data Mining vs Predictive Analytics
Definition: mining discovers patterns/relationships; predictive estimates a specified probable future outcome.
Scenario: products bought together → mining/association; predict next purchase → predictive.
Trap: every model = data mining.
Hook: discover vs forecast.
Source: pp. 481–483.
H — Training vs Validation vs Test
Definition: train fits; validation selects/tunes; test independently judges generalization after choices.
Purpose: prevent the same evidence from both shaping and judging the model.
Scenario: model picked on validation; final report uses untouched test.
Trap: highest training score is proof; repeated tuning on test preserves independence.
Hook: fit, choose, prove.
Source: pp. 488–490, 494–495.
I — MPP vs Distributed File-Based
Definition: MPP partitions data/computation for high-performance parallel analytics; distributed file-based environments flexibly land/store large varied files.
Deciding requirement: interrogative parallel analytical performance vs scalable flexible landing/storage.
Scenario: complex partitionable queries → MPP; huge raw log archive → distributed file.
Trap: one replaces the other universally.
Hook: parallel query engine vs flexible file fabric.
Source: pp. 491–494.
J — Metadata vs Data Quality in Big Data
Metadata: source, structure, lineage, meaning, intended use, quality context.
DQ: validity/completeness/consistency/usability/trust of values for purpose.
Scenario: unknown lineage → Metadata gap; known source with 30% missing key → DQ gap.
Trap: put either discipline at the end.
Hook: know it; trust it.
Source: pp. 472–473, 486–487, 499–500.
K — Predictive vs Prescriptive vs Operational Analytics
Definition: probability/forecast vs action recommendation vs live operational application/response.
Scenario: predict failure → predictive; recommend maintenance → prescriptive; continuously score sensors and shut down unsafe machine → operational.
Trap: treat all three as one maturity ladder. Operational describes use/deployment context.
Hook: likely, action, live.
Source: pp. 481–483, 490–491.
L — Business Value vs Technical Feasibility
Business value: does the problem/outcome justify investment?
Technical feasibility: can data/model/platform/latency/operations deliver and sustain it?
Scenario: model saves $10K/year but needs $2M real-time platform → value/feasibility mismatch.
Trap: impressive prototype automatically deserves production.
Hook: worth doing? can sustain it?
Source: pp. 484–485, 489–491, 496–497.
Final discrimination test
- What changes a default-risk probability into a prescriptive output?
- What facts let you reject “data swamp” for raw files?
- What evidence determines whether 99% training accuracy is meaningful?
- What business fact justifies a speed layer?
- A source is high-quality but has no provenance: Metadata or DQ first?
Answers: recommendation/action; managed context/provenance/DQ; unseen validation/test evidence; actionable latency/value; Metadata.