Module 3 -
From Business
Problems to Data
Solutions
Business Logic
Dr. Tuan Truong
Learning Objectives
Write Build Apply Map Review
Write a Clear Build a Diagnostic Apply the MECE Map to Analytics: Data Analysis
Problem Logic Tree: Break Principle: Ensure Translate "leaf Techniques
Statement: down a high-level your logic is questions" into including
Quantifiable, KPI into root Mutually Exclusive canonical data supervised
scoped, with a causes (driver and Collectively tasks learning,
specific decision branches). Exhaustive. (Classification, unsupervised
owner. Regression, learning, reinforce
Clustering). learning, AI
The BI Value Chain
Data science is not just about code; Example
it is about the flow of value:
Data → Information → Decisions → Actions Data: Raw logs, transactions, sensor data.
Information: Reports, dashboards, patterns.
Decisions: "Should we stop this campaign?" /
"Should we approve this loan?"
Actions: The physical implementation that
impacts the KPI.
If your model creates information but doesn't enable a
Decision, it has zero business value.
What exactly it mean to
be “information”?
Generally, information is
a quantity that reduces
Knowledge is a collection
uncertainty about
of information about
something. So, the better
something.
the information, the more
uncertainty is reduced.
Nature of business problems
➢Data-driven business decision-making problem is
unique, comprising its own combination of goals,
desires, constraints, and even personalities.
➢There are sets of common tasks that underlie the
business problems.
➢The solutions to the subtasks can then be
composed to solve the overall problem.
➢Some of these subtasks are unique to the particular
business problem, but others are common data
mining tasks
Nature of Every data-driven decision
problem is unique because it
mixes:
share a common structure:
• Define the decision
business • Goals (grow revenue, reduce
(“What will we do
differently?”)
cost, improve satisfaction)
problem • Constraints (budget, time, • Define success metrics
(KPIs + thresholds that
capacity, inventory, policies,
compliance) trigger action)
• Trade-offs (speed vs accuracy, • Identify drivers (leading
growth vs margin, automation vs indicators) and root
risk) causes
• Stakeholders & incentives • Select the best action
(finance, marketing, ops, (prioritize, allocate,
leadership; different priorities)
intervene)
Most problems can be decomposed into
subtasks, then recombined:
• Describe = symptoms - what happened
(monitoring, dashboards, segmentation)
Nature of • Diagnose = causes - why it happened
(breakdowns, funnel, cohort, attribution)
business • Predict = prognosis - what will happen
problem (forecasting, churn/propensity models)
• Prescribe = treatment plan what to do
Problem is Domain-specific (optimization, rules, recommendations,
experiments)
Describe what happened
Establish facts and where the change occurred. No theories yet.
Typical outputs
• KPI dashboard: revenue, conversion rate, CAC, AOV, retention, NPS…
• Segmentation cuts: by channel, region, device, product category, customer type (new vs returning)
Example
• Business problem: “Revenue dropped 12% this week.”
• What we should describe:
• Revenue: -12% WoW
• Sessions: flat
• Conversion rate: down from 2.4% → 2.0%
• AOV: unchanged
• Biggest drop segment: Mobile + Paid Social + New customers
• Information: “The drop is mainly conversion, mostly on mobile, coming from paid social, especially new
customers.”
• Identify the driver(s) behind the change—ideally to a small number of
actionable causes.
• Typical tools
• Breakdown analysis: compare before vs after across segments
• Funnel analysis: view → product → add-to-cart → checkout →
payment
• Cohort analysis: behavior by signup/purchase week
• Attribution: shifts in channel mix, campaign quality, or assisted
conversions
• Use the same previous example
Diagnose why • You diagnose with a funnel:
• Product view → Add to cart: stable
it happened • Add to cart → Checkout start: stable
• Checkout start → Payment success: drops sharply on
mobile
• Transaction logs/ops:
• Payment gateway errors on mobile spiked (timeouts)
• Or: new promo code banner covers the “Pay” button on iOS
Safari
• Or: paid social campaign changed targeting to broader
audiences, lowering intent
• Diagnosis statement :
“Revenue fell because mobile checkout completion dropped due
to increased payment failures on iOS (gateway timeout) starting
Monday.”
• Quantify the future impact if nothing changes, and estimate the
upside of fixes.
• Typical outputs
• Forecast: revenue/transactions next 1–4 weeks
• Propensity/churn models: who is likely to churn or buy
• Scenario planning: base / upside / downside depending on
interventions
Predict what • Example (same case)
• Predict
will happen • If mobile payment failure rate stays high, expected
weekly revenue stays -10% to -15% -
• The longer the issue persists, the more new
customers fail to convert and don’t come back,
reducing next month’s retention too
• Predict impact of a fix
• If payment success returns to baseline, revenue likely
recovers ~9–12% (based on historical sensitivity of
revenue to checkout completion)
• Typical tools
• Rules: guardrails/alerts, auto-pausing
campaigns, fraud blocks
• Optimization: budget allocation, pricing,
inventory, routing
• Recommendations: next-best offer,
product recs
Prescribe • Experiments: A/B tests, holdouts,
phased rollouts
Success metrics
• Payment success rate
what to do
• Example:
• Immediate (today)
back to baseline
• Roll back the checkout change OR
switch payment routing to a stable
• Mobile conversion
gateway on mobile recovers from 2.0% to
• Add monitoring + alert: “payment
success rate by device” with
~2.4%
thresholds
• Revenue returns to trend
• Temporarily reduce paid social
spend to mobile until fixed (or line
route to desktop-friendly landing)
• Short-term (this week)
• Run an A/B test: alternative
Decide the best actions (and checkout UI vs current
order), with measurable • Add retry logic + better error
handling
expectations. • Longer-term (this month)
• Build a reliability scorecard for
gateways
• Create an incident playbook: who
responds, what metrics trigger
action, what rollback steps exist
Summary on the example
When you recombine the subtasks, you get a decision narrative like:
• Describe: Conversion fell on mobile paid social (new customers).
• Diagnose: Checkout completion broke due to iOS payment
timeouts.
• Predict: If unresolved, weekly revenue stays -10% to -15% and
retention damage increases.
• Prescribe: Roll back + reroute payments now, add alerts, then
experiment with checkout improvements.
Business problems vs analytics tasks
Why each problem is unique (context) Common analytics tasks
1) Different goals A) Describe (What happened?)
• Grow revenue / profit • KPI reporting, dashboards
• Reduce cost / risk • Segmentation, cohort, funnel
• Improve experience / speed
B) Diagnose (Why did it happen?)
2) Real constraints • Breakdowns by channel/region/device
• Budget, time, people, capacity • Root-cause drill-down, attribution
• Inventory, supply, SLA
• Policy, compliance, privacy
C) Predict (What will happen next?)
3) Trade-offs everywhere • Forecasting demand/sales
• Growth vs margin • Propensity/churn/fraud prediction
• Speed vs accuracy
• Automation vs control D) Prescribe (What should we do?)
• Budget allocation, prioritization
4) Stakeholders & incentives • Optimization, rules, recommendations
• Marketing wants growth • Experiments (A/B tests) to confirm impact
• Finance wants efficiency
• Ops wants stability E) Close the loop
→ alignment is part of the problem • Action → measure → learn → improve
3-layer question
Decision question (business view)
→ Analytical question (analysis view)
→ Data question (data/engineering view)
Decision question
• a question meant to drive an action/choice
(yes/no, choose A vs B, increase/decrease,
prioritize what).
• Answer decision question to take a
concrete action with a clear goal/KPI and
scope.
• Template:
• “Should we [DO ACTION A] instead of
[B / current approach] for
[SEGMENT/TARGET] in [TIME /
CHANNEL] to improve [TARGET KPI]
(expected +Δ) while keeping
[GUARDRAILS] within limits?”
Decision question
• 6 parts
• Decision-maker + controllable lever (what your team
can actually change)
• Alternatives (A/B/…): how much discount? how much
budget increase? what process change?
• Goal: the KPI / north-star metric to improve (revenue,
profit, retention…)
• Constraints / guardrails: NPS, refunds, churn,
complaints, stockouts… (avoid optimizing the wrong
thing)
• Scope: segment, channel, geography, product line
• Timeframe: which week/month/quarter, and how long
you’ll measure
• avoiding questions that are too technical or too vague.
• Too vague: “How do we increase revenue?” → no lever,
no scope, no timeframe.
• Too technical: “Should we use XGBoost or LSTM?” →
jumping to solution before clarifying the decision.
Analytical
questions
and Data
questions
Decision (what to do?)
→ Analysis (what must we measure/prove to decide?)
→ Data (what data is needed to do the analysis?)
Data Questions(data needs, grain, keys, joins)
To specify:
• Entity & grain: what is one row? (one order, one session, one
user-day…)
• Required event logs: view/click/add_to_cart/order/refund…
• Dimensions: channel, campaign, geo, product, cohort…
• Time window + metric definitions: how is 7-day retention
defined? net vs gross revenue?
• Join keys: user_id, order_id, session_id…
• Data quality checks: missing values, duplicates, timezone…
Data question template
• “Which tables, fields, join keys, and time range do we need to
compute [metric] and break it down by [dimensions]?”
The Problem Statement
Clinic
The "Golden Template"
• "In [Time Window], for [Segment/Scope], the [KPI] changed
from [Baseline] to [Current].
• We need to decide [Decision] to reach [Target] while
respecting [Constraints].
• Decision Owner: [Role]"
The "Sanity Check" Checklist:
• [ ] Is there a specific Number (Metric)?
• [ ] Is the Timeframe defined?
• [ ] Is the Scope clear (not "everyone")?
• [ ] Is there an actionable Decision?
Good vs Bad Problem Statement
The Vague Statement The Quantifiable Statement
"Our sales are dropping, and we need to use AI "In Q3 2024, for iOS Mobile Users, Cart
to find out why and fix it.“ Conversion Rate dropped from 5.0% to 3.2%.
We need to decide which checkout steps to
simplify to return to 4.5% within 3 months.
Owner: Product Manager."
Why the fails: No baseline, no target, no specific scope, implies "AI" is the goal (it's a tool, not a goal).
Logic Trees
• Breaking Down the Problem
• We cannot solve "Profit is down." We must
find the Root Cause.
Remember the Root-cause analysis with 5-whys and Fish-bone diagram
• Structure of a Diagnostic Tree:
• Root: The Symptom (e.g., Profit ↓).
• Branch Level 1: Mathematical
breakdown (Revenue ↓ OR Cost ↑).
• Branch Level 2: Drivers (Revenue = Traffic
× Conversion × Price).
• Leaves: Testable Hypotheses /
Questions.
In-class activity - Build the Tree
Group Exercise
1 2 3
Scenario: Choose a problem Perform Root-cause analysis Complete the tree:
(e.g., "Student enrollment with 5-whys and Fish-bone • Build 1 Diagnostic Issue Tree (Why
has dropped 15%"). diagrams is this happening?).
• Drill down at least 3 levels deep.
• Identify 6–10 Leaf Questions (What
would we measure?).
• Select your Top 3 Hypotheses
(Ranked by likely impact).
The MECE Principle
Structuring for Completeness
MECE: Mutually Exclusive, Collectively Exhaustive.
Mutually Exclusive (No Overlap):
• Bad: "Customers under 30" vs. "Students". (A student can be
under 30).
• Good: "Age < 20" vs. "Age 20-30" vs. "Age > 30".
Collectively Exhaustive (No Gaps):
• Bad: "Facebook Traffic" vs. "Google Traffic". (What about
Direct? Email? Referral?).
• Good: "Paid Social", "Organic Search", "Direct", "All Others".
Audit Your Tree
Before moving to data, ask:
No Overlap: Can a single data point belong to two branches?
No Gaps: Is there a major category missing? (Use an "Other"
bucket if needed).
Parallel Logic: Are you mixing "Strategic levels" with
"Operational bugs" at the same hierarchy?
Testability: Can every leaf be validated with data?
Mapping to Analytics
• From Business to Data Science
• We must convert Leaf Questions into Canonical Tasks.
Business Question (Leaf) Canonical Task Type
"Are there distinct groups of
Clustering Unsupervised
users?"
"Which users will likely leave
Classification Supervised
next?"
"How much will sales be next
Regression Supervised
month?"
"What items are bought
Association Unsupervised
together?"
Data analytics review
It’s not about how you analyze data but how we apply the right technique in
the right scenario
SUPERVISED VS UNSUPERVISED LEARNING
➢Supervised method learn
about the target
information through the
set of examples.
given a specific purpose for the
grouping—predicting the target
there must be data on the target
in the dataset
The value for the target variable
for an individual is often called
the individual’s label
➢An unsupervised learning
task might involve the
same set of examples but
would not include the
target information
➢Classification and class probability estimation attempt
to predict, for each individual in a population, which of a
(small) set of classes this individual belongs to.
(classification model and scoring model)
Example: “Among all the customers of Mobiphone,
which are likely to respond to a given offer?”.
Two class: will and will not. -> create a classification
model for determine which class a customer is belong
to.
Evaluate each individual customer and produce a score
of how likely each is to respond to the offer -> create a
scoring model: applied to an individual produces,
instead of a class prediction, a score representing the
probability that that individual belongs to each class.
➢Regression (“value estimation”) attempts to estimate or
predict, for each individual, the numerical value of some
variable for that individual.
Example: “How much will a given customer use the
service?” The property (variable) to be predicted here is
service usage, and a model could be generated by
looking at other, similar individuals in the population
and their historical usage.
➢Causal modeling attempts to help us understand what
events or actions actually influence others.
Example: consider that we use predictive modeling to
SUPERVISED LEARNING target advertisements to consumers, and we observe
that indeed the targeted consumers purchase at a
higher rate subsequent to having been targeted.
ALGORITHMS
Was this because the advertisements influenced the
consumers to purchase?
Or did the predictive models simply do a good job of
identifying those consumers who would have
purchased anyway?
CLASSIFICATION ALGORITHMS
Classification is the process of
recognizing, understanding, and
grouping ideas and objects into
preset categories or “sub-
populations.” Using pre-
categorized training datasets,
machine learning programs use
a variety of algorithms to
classify future datasets into
categories.
Read more
REGRESSION (VALUE ESTIMATION)
The future values are predicted with the help
of regression algorithms in Machine
Learning. The input data/historical data is
used to predict a wide range of future values
using regression. Label in ML is defined as
the target variable (to be predicted) and
regression helps in defining the relationship
between label and data points. Regression is
a type of supervised learning in ML that
helps in mapping a predictive relationship
between labels and data points.
Popular regression algorithms
CASUAL MODELING
Famous algorithms:
- Structural Equation Modeling(SEM)
- Fast causal inference - FCI
- Greedy Equivalence Search - FGES
Constraint-based algorithms
Score-based algorithms
Methods and tools for causal
discovery
Read more
OVERFITTING VS UNDERFITTING MODEL
Problem Overfitting Underfitting
explain gives accurate predictions for training data but not a scenario in data science where a data
for new data model is unable to capture the
relationship between the input and
output variables accurately, generating a
high error rate on both the training set
and unseen data.
Reasons • The training data size is too small and does not not trained for the appropriate length of
contain enough data samples to accurately time on a large number of data points.
represent all possible input data values.
• The training data contains large amounts of
irrelevant information, called noisy data.
• The model trains for too long on a single sample
set of data.
• The model complexity is high, so it learns the
noise within the training data.
Test • K-fold cross-validation
divide the training set into K equally sized subsets
or sample sets called folds
➢Clustering attempts to group individuals in a
population together by their similarity, but not
driven by any specific purpose. Clustering is
useful in preliminary domain exploration to see
which natural groups exist because these
groups in turn may suggest other data mining
tasks or approaches. Usually used in decision-
making processes
➢Example: “Do our customers form natural groups or
segments?”
➢Co-occurrence grouping (also known as
frequent itemset mining, association rule
discovery, and market-basket analysis)
attempts to find associations between entities
based on transactions involving them.
▪ Example: What items are commonly purchase
together?
➢Profiling (also known as behavior description)
attempts to characterize the typical behavior
UNSUPERVISED LEARNING
of an individual, group, or population
Example: “What is the typical cell phone usage of
this customer segment?” Behavior may not have a
ALGORITHMS
simple description; profiling cell phone usage might
require a complex description of night and weekend
airtime averages, international usage, roaming
charges, text minutes, and so on.
CLUSTERING & CO-OCCURRENCE GROUPING
4 Types of clustering:
- Hierarchy
- Centroid
- Density-based
- Distribution-based
8 Clustering Algorithms
FOR BOTH LEARNING
➢Similarity matching attempts to identify similar individuals based on data known
about them. Similarity matching is the basis for one of the most popular methods for
making product recommendations. Similarity measures underlie certain solutions to
other data mining tasks, such as classification, regression, and clustering.
Example: P&G is interested in finding companies similar to their best business customers, in order to focus
their sales force on the best opportunities. They use similarity matching based on “firmographic” data
describing characteristics of the companies.
➢Link prediction attempts to predict connections between data items, usually by
suggesting that a link should exist, and possibly also estimating the strength of the
link.
➢Data reduction attempts to take a large set of data and replace it with a smaller set
of data that contains much of the important information in the larger set. The smaller
dataset may be easier to deal with or to process. Data reduction usually involves loss
of information. What is important is the trade-off for improved insight.
MACHINE LEARNING
VS DEEP LEARNING
➢Machine Learning
comprises of the ability of the
machine to learn from trained
data set and predict the
outcome automatically
➢Deep learning can
automatically find out the
attributes from raw data while
machine learning selects
these features manually
which further needs
processing
employs artificial neural networks
with many hidden layers, big data,
and high computer resources
REINFORCEMENT LEARNING
RL is learning what to do, from real situations to give
certain actions, as long as maximize the reward.
not told about the action to take, but instead must
discover the action that can generate the most rewards.
a concept called an agent, it has a bit of an implication
about an entity that you want to train to be able to do a
certain task that you entrust.
Deep Q-Learning is a model-free reinforcement learning
algorithm. It learns based on values. Value-based
algorithms update the value function based on an
equation (especially the Bellman equation). While the
other, policy-based, estimates the value function with a
greedy policy resulting from the last policy improvement.
[Link]
djeZ1GEY5Wz
[Link]
deep-q-learning-choi-game-cartpole-Az45bYy6lxY
GENERATIVE AI
•trains a model to learn the data distribution and generate new
content that looks like the training data.
•Instead of predicting a label (e.g., Car / Not Car), it produces
new samples (e.g., text, images, audio, code).
•differs from “traditional” ML (Discriminative AI)
•Discriminative models: learn a boundary to classify/predict
Example: “Is this email spam?”
•Generative models: learn how data is formed to
create/simulate. Example: “Write an email in this style.”
Common model families Practical applications
•Autoregressive models (e.g., Transformers): generate next token •Text: summarization, translation, tutoring, customer
step-by-step support drafts
•Diffusion models: start from noise, iteratively denoise to create •Images: product mockups, design variations,
images marketing creatives
•GANs: generator vs discriminator competition to create realistic •Audio: voice synthesis, music generation, noise
samples removal
•VAEs: learn a compressed latent space and sample from it •Code: code completion, unit tests, documentation
generation
HOMEWORK
Individual: Finish and submit BPMN for business process
Group project: Stage 1 - Problem framing & Logic tree