[ML02] CRISP-DM, KPI Tree & Problem Framing.
Updated Jul 20, 2026

Tags: machine learning data science business analytics

Bài 2: CRISP-DM, KPI Tree & Problem Framing

Series: Business Analytics Cheatsheet Series

Topics: crisp-dm kpi-tree problem-framing ab-testing guardrails

Source: BA1 — Introduction to Business Analytics: CRISP-DM · KPI Tree · Problem Framing. Organized by CRISP-DM phase. Cross-phase topics (KPI Tree, problem framing, templates) sit in their own sections at the end.

Core claim of the deck: up to 80% of analytics failures come from a misframed problem or KPI — not from a weak model. Framing is the highest-leverage step.


0. Orientation

Term Type Definition / Usage
Business Analytics Discipline Process that turns data into insight and action. Spans descriptive → diagnostic → predictive → prescriptive layers. Outputs: reports/KPIs, models, A/B tests, operational recommendations.
BI vs DA/DS vs AI/ML Taxonomy BI: reporting, dashboards, KPI monitoring (past/present). DA/DS: root-cause, forecasting, segmentation, optimization. AI/ML: learn from data to automate prediction/recommendation inside the product.
Data Value Chain Framework Collect → Store → Clean/Transform → Analyze/Model → Deploy → Monitor. Treat as a product lifecycle: quality, reliability, repeatability. Principle: start from the decision to be supported, not from the data you happen to have.
Team roles Org Product/Data PM, Data Analyst, Data Scientist, Data Engineer, BI Dev, MLOps/Platform. Core skills: SQL, statistics, EDA, visualization, modeling, data storytelling. Supporting: domain knowledge, experiment design, data ethics & governance.
Trustworthy AI Governance Fairness, safety, transparency, auditability. Operationalized through KPI definitions, data quality, lineage & catalog, RACI. Privacy/security/compliance are constraints, not afterthoughts.
CRISP-DM Framework Cross-Industry Standard Process for Data Mining. Six phases, explicitly iterative — loop back when new insight or a goal change appears. Artifacts per phase: charter, data dictionary, feature list, model card, runbook.

1. Business Understanding (BU)

Goal: convert a business objective into a decision, a measurable success definition, and a bounded scope.

Term Type Definition / Usage
SMART objective Technique Specific, Measurable, Achievable, Relevant, Time-bound. Example: “+15% Q4 revenue YoY.” Vague goals (“improve engagement”) cannot be evaluated and should be rejected at charter review.
Decision-centric framing Principle Identify the decision analytics will inform before choosing a method. Test question: “What decision will this inform?” If there is no answer, do not start the work.
Success KPI Metric The primary measure the project is judged on. One primary; 2–4 supporting. More than one primary = no primary.
Guardrail Metric Hard constraint. Example: “do not reduce overall CR by more than 1pp.” Any solution violating it is not deployed — state this in the charter, not in the readout.
Hypothesis (H1, H2…) Artifact Testable statement linking a lever to a KPI. Example: H1 = bundles/upsell raise AOV without hurting CR; H2 = mobile checkout optimization raises mobile CR. Each must name the data needed to test it.
Project Charter Artifact Business problem & opportunity, objective, in/out of scope, stakeholders & RACI, hypotheses, success criteria & guardrails, timeline, risks & mitigation. The single approval gate before data work starts.
Scope (In / Out) Artifact Explicit exclusions matter more than inclusions. Example out-of-scope: paid acquisition, pricing/discount policy, backend overhaul, desktop-only work.
RACI Governance Assign per workstream: Sponsor = Accountable; Marketing/UX/DS = Responsible; DE/Finance = Consulted; Exec = Informed.
Risk register Artifact Risk → mitigation pairs. Typical: solution hurts CR → guardrail monitoring + phased A/B + rollback; poor data quality → early audit + rule-based fallback; Q4 seasonality → use historical + current-year data.
Data rules / canonical definitions Governance Agree on what a session, transaction, user, and attribution are before measuring. Disagreement here invalidates everything downstream.

BU outputs checklist: problem statement · scope · stakeholders & RACI · analysis questions · hypotheses · KPIs & guardrails · timeline · risks · data definitions.


2. Data Understanding (DU)

Goal: verify that reality matches the charter’s assumptions before committing engineering time.

Term Type Definition / Usage
Source inventory Technique For each source: schema, size, freshness, access permissions. Gate the project on access being granted, not promised.
EDA (at DU stage) Technique Distributions, outliers, missingness, temporal consistency. Purpose here is validation, not insight generation.
KPI reconciliation Technique Match KPI definitions against real fields: order date vs ship date, returns, cancellations, order edits. Expect a gap; document it.
Funnel analysis Technique session → view → add_to_cart → checkout → purchase. Quantify absolute drop-off per stage, not just rates.
Segmentation (quick win) Technique Split by channel/device (mobile vs desktop, organic vs paid) before modeling. Often reveals that the “average” problem does not exist in any segment.
Double-counting detection Diagnostic Retries and timeouts inflate purchase counts. Deduplicate on transaction key before any revenue claim.
Leakage detection (early) Diagnostic Flag fields generated after the decision point at DU, not at modeling. Cheaper to exclude than to debug later.

3. Data Preparation (DP)

Goal: produce a leakage-safe, reproducible modeling dataset.

Term Type Definition / Usage
Cleaning Technique Handle missing values and outliers; normalize units, currency, timezone. Do this inside a fitted pipeline, not in an ad-hoc notebook cell.
Feature engineering Technique Derive decision-relevant signals: RFM (recency/frequency/monetary), price_sensitivity, device_speed. Domain knowledge beats feature count.
Time-based split Technique Train on pre-T0, evaluate on post-T0. Mandatory whenever the production use is forecasting a future event. Random splits leak temporal information.
Data tests Technique Schema, ranges, freshness, uniqueness. Run as pre-deploy gates with alert thresholds.
Taxonomy & data dictionary Artifact Standardized event naming, field semantics, late-arriving data policy.
Feature Store Infrastructure Reuse features across models with lineage and versioning; guarantees train/serve parity.
Data leakage Anti-pattern Any information unavailable at decision time entering training. Symptom: excellent validation, collapse in production. See file 03 for the full taxonomy.

4. Modeling (M)

Goal: the simplest model that supports the decision, with a defensible baseline.

Term Type Definition / Usage
Task selection Decision Classification (churn) · Regression (AOV) · Clustering (segments) · Optimization (pricing). Derived from the decision, not from data availability.
Baseline Practice Simple model + business rules first. Compare AUC/MAE/lift against it. A model that does not beat rules should not ship.
Threshold tuning Technique Set the decision threshold from business goals and error costs, not from the 0.5 default. See file 06.
Time-aware cross-validation Technique Rolling-origin / TimeSeriesSplit. Standard k-fold overstates performance for temporal targets.
Regularization & feature selection Technique L1/L2 and embedded importance to control overfitting and reduce serving cost.
Explainability (e.g. SHAP) Technique Attribute predictions to features. Required for regulated decisions and for stakeholder trust; treat as an explanation, not a causal statement.
Bias check Governance Compare performance across protected/vulnerable segments before deployment.
Packaging Practice Model card, reproducible environment, dependency locking, seed fixing.

5. Evaluation (E)

Goal: separate technical performance from business impact — they are not the same and often disagree.

Term Type Definition / Usage
Technical evaluation Metric set AUC/PR-AUC, RMSE/MAE, calibration, cohort stability. Answers: does the model rank/predict well?
Business evaluation Metric set Revenue/profit uplift, KPI movement, guardrail impact. Answers: does deploying it make money without breaking anything?
Experimentation Method A/B test, holdout, careful sequential testing. The only credible bridge from model metric to business metric.
Cohort stability Diagnostic Performance consistent across cohorts and time slices. Instability predicts production degradation.
“Technical success, business failure” Anti-pattern High AUC, zero adoption. Fix by involving the decision owner from BU and by testing the intervention, not the model.

6. Deployment (D)

Goal: operate the model as a product with defined failure modes.

Term Type Definition / Usage
Batch vs real-time scoring Decision Batch for periodic campaigns; real-time when the decision happens in a user session. Latency budget belongs in the charter.
Canary / Blue-Green Deployment pattern Canary = gradual traffic ramp; Blue-Green = atomic switch between two full environments. Both require an automated rollback trigger.
Monitoring Practice Data drift, model drift, SLA/latency, alerting, auto-rollback. Silence is not health — instrument the absence of predictions too.
MLOps Practice CI/CD, model registry, feature–model consistency.
Handover package Artifact Project charter · data dictionary · EDA report & quality checklist · model card · evaluation report · rollout plan · monitoring dashboard · operations runbook · improvement plan.

CRISP-DM common pitfalls

Pitfall Type Fix
Starting from data, not decisions Anti-pattern Ask “what decision does this inform?” at charter review; kill the project if unanswered.
Unmeasurable KPIs Anti-pattern Require formula + window + unit + owner for every KPI.
Data leakage Anti-pattern Time-based splits; fit preprocessing on training folds only; exclude post-decision fields.
Inconsistent KPI definitions across teams Anti-pattern One canonical definition in the data dictionary; dashboards read from it.
Great model, zero business impact Anti-pattern Design the intervention and the experiment alongside the model, not after it.

KPI Tree

Purpose: connect a North Star to levers a named person can pull.

Term Type Definition / Usage
KPI Tree Framework Hierarchical decomposition: North Star → drivers → levers → owners. Makes trade-offs explicit and prevents disconnected metric sprawl.
5-step process Process 1) Choose the North Star (Monthly Revenue, MRR, GMV). 2) Decompose into influenceable drivers. 3) Attach levers and owners per branch. 4) Define guardrails, trade-offs, instrumentation, update cadence. 5) Connect to dashboards + weekly review.
KPI selection principles Principle Fewer but better; measurable, available, actionable. Avoid vanity metrics unless tied to a decision.
E-commerce revenue tree Example Revenue = Sessions × CR × AOV. CR = P(view→cart) × P(cart→checkout) × P(checkout→purchase). AOV = Σ(price × qty)/#orders. Levers: bundles, upsell, free-shipping threshold.
Numeric illustration Example 1,000,000 sessions/mo · CR 2.0% · AOV 420k → revenue ≈ 8.4bn VND. Target +15% → 9.66bn. O1: CR → 2.3% (AOV fixed). O2: AOV → 483k (CR fixed). Quantifying options this way turns strategy debate into arithmetic.
SaaS MRR tree Example MRR = #Customers × ARPA × (1 − Churn); #Customers = New Signups × Activation Rate × Conversion to Paid. Levers: pricing tiers, onboarding, success motions, retention plays.
Measurement quality Practice Canonical definitions (returns/cancellations, order edit dates); fixed windows for CR/AOV by channel/device; bot traffic removed. Health metrics alongside: latency, payment errors, inventory availability.
Measurement plan / instrumentation Artifact Events: page_view, add_to_cart, begin_checkout, purchase. Schema: user_id, session_id, device, source, campaign, value. Fixed update frequency and named dashboard owner.
Dashboard design Practice Board 1: North Star + driver-tree heatmap (R/Y/G). Board 2: funnel by device/channel. Board 3: AOV by category. Features: cohort filters, event annotations (sales, UI changes).
KPI Tree pitfalls Anti-pattern Overlapping branches / double counting · missing guardrails · no clear owner or lever · weak causal validation · refresh slower than the decision cadence.

Problem Framing

Term Type Definition / Usage
DOC — Decision, Options, Criteria Framework D: the choice to be made. O: candidate options. C: the KPI set used to compare them. A problem statement without all three is not actionable.
Problem Statement Template Artifact Context (current situation + goal) → Decision → Options → Criteria & KPIs. One page maximum.
Business question → analytics task Technique “Grow revenue 15% without more ad spend” → Task 1 (diagnostic): find funnel bottlenecks & affected cohorts. Task 2 (prescriptive): rank improvements by ROI. Always translate before choosing a method.
Target variable definition Technique Make it explicit and time-bounded: “purchase within 7 days of session (binary).” Ambiguous targets produce unusable models.
Decision-point leakage Anti-pattern Excluding variables generated after the decision point (e.g. applied voucher). Pair with time-based splits and drift/cohort monitoring.
Constraints / Assumptions / Risks Artifact Constraints: engineering resources, rollout time, SLA. Assumptions: stable seasonality, data reliability. Risks: ad policy change, supply disruption, behavior shift. State assumptions explicitly so they can be invalidated.
Success criteria & guardrails Artifact Primary: +15% revenue. Secondary: CR drop <1pp, stable AOV. Technical: lift ≥ X%, acceptable SE/CI, latency <200ms. Ethical: no worsened experience for vulnerable groups; transparent pricing.
Impact/Effort matrix Prioritization Rank opportunities into quick wins · big bets · fillers · avoid. Output: a 6–8 week backlog with milestones.

Anti-patterns & fixes

Anti-pattern Fix
Analytics not tied to a decision Ask “what decision will this inform?” before scoping.
Vague / unmeasurable KPIs Standardize definitions and sources in the data dictionary.
Technical success, business failure Experiment; canary release; measure the intervention.

A/B Testing basics

Term Type Definition / Usage
Randomization unit Design choice User or session. Choose the unit the intervention actually affects; mismatches cause interference and inflated significance.
Primary metric Design choice One metric, declared before launch. Everything else is secondary or a guardrail.
Peeking Anti-pattern Checking results repeatedly and stopping at significance. Inflates false positives; fix the duration in advance or use a sequential test designed for it.
Group independence Assumption No spillover between arms. Violated by social features, shared inventory, marketplace supply.
Test duration Design choice Derived from traffic and minimum detectable effect (MDE), plus at least one full weekly cycle.
Sample size rule of thumb (CR) Formula n ≈ 16 × p(1−p) / Δ² per group. Example: p = 2%, Δ = 0.3pp (0.003) → n ≈ 34,844 per group. Use a proper calculator for multiple KPIs or adjusted tests.

Templates

Problem Framing Canvas

Context & Goals:
Decision (D):
Options (O):        O1 …  O2 …  O3 …
Criteria & KPIs (C):
Constraints:
Assumptions:
Risks:
Experiment / Measurement plan:
Owner:              Timeline:
Required data preparation:

KPI Tree (blank)

North Star: ______________________________
  Branch 1 → driver → lever → owner
  Branch 2 → driver → lever → owner
  Branch 3 → driver → lever → owner
Guardrails: ______________________________
Instrumentation & cadence: _______________

CRISP-DM checklist

BU: SMART goals · KPIs/guardrails · stakeholders & RACI
DU: inventory · EDA · data definitions · quality issues
DP: missing/outliers · feature list · time-aware splits
M : baselines · evaluation · explainability · versioning
E : business evaluation · experiment plan
D : deployment · monitoring · runbook · rollback

Definition of Done

References: Davenport & Harris, Competing on Analytics · Provost & Fawcett, Data Science for Business · CRISP-DM guide · Google Analytics / Amplitude event taxonomy docs.


🔗 Cùng series

Bài 1: Business Analytics Foundations & Types of Analytics
Bài 3: Data Quality & Feature Engineering


Nếu bài này hữu ích, hãy chia sẻ cho người đang học Data Science / Business Analytics. Mọi góp ý về lỗi kỹ thuật rất được hoan nghênh.