Overview
Applied Data Science has three topics. Bringing data science to life in Web3 looks at blockchain data: public, append-only and pseudonymous, and what that means for analysis. Big data case studies apply the course to consumer goods and services, technology and healthcare, plus a case on the Enterprise Big Data Framework Alliance. Data ethics covers privacy and transparency, bias and fairness in AI, and accountability.
The module integrates everything before it. A case study asks you to frame the problem as in Module 1, choose methods from Modules 2 to 5 and 8, scale them as in Module 7, and present them as in Module 6. Ethics then asks whether the result is lawful, fair and explainable. On the exam, this is where a technically correct answer can still be wrong because it ignores privacy, bias or ownership.
How it's tested
Expect longer case-style scenarios with BEST, MOST appropriate or FIRST phrasing. Several options will be technically workable; the right one also respects privacy law, fairness and accountability, and fits the industry context. Read for the constraint the scenario plants, such as sensitive health data, a protected group or a public ledger.
-
Apply the toolkit to an industry
A consumer goods, technology or healthcare problem is described and you choose the method, data approach and success measure that fit the domain.
-
Find the ethical flaw
A model or data practice is described and you identify the privacy, bias or transparency issue, such as a proxy variable or re-identification risk.
-
Assign responsibility
Something goes wrong after deployment and you choose the accountability mechanism, such as an owner, documentation, an audit trail or human review.
A hospital wants to share patient data with a research partner and proposes replacing names with random IDs. Pseudonymized data is still personal data under GDPR because it can be re-linked, and remaining fields such as birth date, sex and postal code can re-identify people. The better answer applies data minimisation, stronger de-identification such as aggregation or k-anonymity, and a lawful basis with an agreement, not ID replacement alone.
Topics 3
Official topics from the Enterprise Big Data Framework Alliance outline, in outline order.
9.1 Bringing Data Science to Life in Web3Analyzing blockchain and decentralized application data, and its specific opportunities and constraints.
- What it means
- Web3 refers to applications built on public blockchains, where transactions and smart-contract activity are recorded on a shared, append-only ledger. That creates large, openly accessible datasets of transactions, wallets and tokens that data scientists can analyze, typically through nodes, indexing services or exported datasets.
- What to understand
- On-chain data is transparent and pseudonymous: addresses are visible, names are not, but clustering and linkage can de-anonymize users. Common uses include fraud and anomaly detection, wallet clustering, token and transaction analytics. Immutability conflicts with privacy rights such as erasure, so personal data generally should not be written on-chain. Off-chain data enters smart contracts through oracles, which become trusted points of failure.
- How it shows up
- A team plans to analyze wallet transactions and calls the data anonymous, and you identify linkage-based re-identification risk.
- Common confusion
- Equating pseudonymous with anonymous; a public, permanent transaction history can often be tied to real identities.
9.2 Big Data Case StudiesApplying the EBDS toolkit in consumer goods and services, technology, healthcare and the Enterprise Big Data Framework Alliance.
- What it means
- Case studies show how the same methods change with context. Consumer goods and services use demand forecasting, recommendation, market-basket analysis and churn models; technology firms use large-scale experimentation, recommendation and anomaly detection; healthcare uses risk prediction and medical imaging under strict privacy and validation demands.
- What to understand
- Domain constraints drive method choice: imbalanced outcomes and high false-negative costs in healthcare favor recall and clinical validation; huge user bases in technology favor controlled A/B tests over observational comparisons; retail seasonality calls for time-aware validation. Each case still follows the same process: frame, prepare, model, evaluate, communicate and govern.
- How it shows up
- A hospital readmission model shows high accuracy on imbalanced data and you choose recall-focused evaluation and clinical review, or a tech firm wants to prove a feature's effect and you choose a randomized experiment.
- Common confusion
- Transferring a method that worked in one industry without adjusting for that industry's data, risks and regulation.
9.3 Data EthicsPrivacy and transparency, bias and fairness in AI, and accountability.
- What it means
- Data ethics asks whether data use is lawful, fair, transparent and owned. Privacy covers collection, consent, minimisation and de-identification; bias and fairness cover how models can disadvantage groups; accountability covers who answers for a model's outcomes and how decisions are documented and reviewed.
- What to understand
- Under GDPR, pseudonymized data is still personal data, and principles include purpose limitation and data minimisation. Bias enters through historical, sampling, measurement and label bias, and through proxies such as postal code for ethnicity; dropping the protected attribute does not remove it. Fairness metrics such as demographic parity, equal opportunity and equalized odds can conflict when base rates differ. Accountability uses named owners, model documentation, audit trails, impact assessments and human oversight.
- How it shows up
- A credit model excludes race and gender but approval rates still differ by race, and you identify proxy bias and the need to measure outcomes by group.
- Common confusion
- Believing fairness through unawareness works: omitting a protected attribute does not stop correlated features from reproducing the bias.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Pseudonymous vs. anonymous data
Pseudonymous data replaces identifiers but can be re-linked; anonymous data cannot reasonably be linked to a person.
- Remember
- Under GDPR, pseudonymized data remains personal data; truly anonymized data falls outside its scope.
- Tested as
- Data-sharing and Web3 scenarios.
- Trap
- Treating random IDs or wallet addresses as anonymous.
-
Quasi-identifiers and re-identification
Fields such as birth date, sex and postal code are not identifying alone but can identify people in combination.
- Remember
- Removing names is not enough; generalize, aggregate or apply k-anonymity to quasi-identifiers.
- Tested as
- Privacy risk questions in health and consumer cases.
- Trap
- Releasing detailed records after dropping only names and ID numbers.
-
Data minimisation and purpose limitation
Collect only the data needed for a specified purpose and do not reuse it for incompatible purposes.
- Remember
- More data is not automatically better if it lacks a lawful, stated purpose.
- Tested as
- Scenarios proposing to collect or reuse data broadly.
- Trap
- Choosing to keep all data in case it becomes useful later.
-
Differential privacy
Adds calibrated noise to query results or training so that any one individual's presence has a limited effect on outputs.
- Remember
- It trades some accuracy for a formal privacy guarantee controlled by a privacy budget (epsilon).
- Tested as
- Options for sharing statistics from sensitive data.
- Trap
- Expecting full accuracy with strong privacy protection.
-
Sources of bias
Historical bias (past inequities in the data), sampling or representation bias (groups under-represented), measurement bias (flawed proxies for the target) and label bias (biased human labels).
- Remember
- Identify where bias enters before choosing a fix.
- Tested as
- Root-cause questions about unfair models.
- Trap
- Assuming a larger dataset removes historical bias.
-
Proxy variables
Features correlated with a protected attribute, such as postal code with ethnicity, that let a model reproduce discrimination indirectly.
- Remember
- Fairness through unawareness fails when proxies exist; test outcomes by group.
- Tested as
- Credit, hiring and insurance scenarios.
- Trap
- Declaring a model fair because the protected attribute was removed.
-
Fairness metrics
Demographic parity compares positive-outcome rates across groups, equal opportunity compares true positive rates, and equalized odds compares both true and false positive rates.
- Remember
- When base rates differ, several fairness criteria cannot all be met at once, so the choice must be justified for the context.
- Tested as
- Questions on how to measure fairness.
- Trap
- Assuming one universal fairness metric exists.
-
Transparency and explainability
People affected by automated decisions should be able to understand how decisions are made, and GDPR gives rights around solely automated decisions with significant effects.
- Remember
- High-stakes decisions call for explainable models or explanation methods plus meaningful human review.
- Tested as
- Model choice in regulated or high-impact cases.
- Trap
- Choosing the most accurate black box for a decision that must be justified to individuals.
-
Accountability mechanisms
Named model owners, documentation such as model cards and datasheets, audit trails, impact assessments and human oversight.
- Remember
- Accountability is assigned before deployment, not discovered after an incident.
- Tested as
- Post-incident and governance questions.
- Trap
- Blaming the algorithm as if no one owned its use.
-
Immutability vs. privacy
Blockchain records cannot be changed or deleted, which conflicts with rights to rectification and erasure.
- Remember
- Keep personal data off-chain and store only references or hashes where necessary.
- Tested as
- Web3 design scenarios.
- Trap
- Writing personal data to a public chain for transparency.
-
Randomized experiments (A/B tests)
Randomly assign users to control and treatment to estimate a change's causal effect.
- Remember
- Randomization balances confounders that observational comparisons cannot.
- Tested as
- Technology case studies proving a feature's impact.
- Trap
- Comparing users who opted in with those who did not and calling the difference causal.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Anonymization vs. pseudonymization vs. differential privacy
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Anonymization | Irreversibly removing the ability to identify individuals | Outside GDPR scope if truly achieved; hard to guarantee | Claiming anonymity after removing names only |
| Pseudonymization | Replacing identifiers with tokens that can be re-linked with extra information | Still personal data; reduces risk but not obligations | Sharing pseudonymized data as if it were anonymous |
| Differential privacy | Calibrated noise limits what outputs reveal about any individual | Formal, tunable guarantee at some cost in accuracy | Ignoring the accuracy cost of a small privacy budget |
Demographic parity vs. equal opportunity vs. equalized odds
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Demographic parity | Equal rates of positive outcomes across groups | Ignores whether outcomes were deserved | Applying it where true base rates legitimately differ |
| Equal opportunity | Equal true positive rates across groups | Qualified people have equal chances of approval | Overlooking unequal false positive rates |
| Equalized odds | Equal true positive and false positive rates across groups | Stricter; both error types balanced | Expecting to satisfy it together with calibration when base rates differ |
Historical vs. sampling vs. measurement bias
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Historical bias | Data reflects past unfair decisions or inequalities | Present even with perfect sampling | Thinking more of the same data fixes it |
| Sampling bias | Some groups are under- or over-represented | Fixed by better data collection or reweighting | Assuming large volume means representative |
| Measurement bias | The measured target is a flawed proxy, such as healthcare spending for illness | The label itself is skewed | Trusting a convenient proxy target without checking it |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Fairness through unawareness
- What candidates pick
- Accept a lending model as fair because race and gender were removed.
- Why it looks right
- The model cannot see the protected attributes directly.
- Why it's wrong
- Correlated features such as postal code act as proxies, so disparities persist; fairness must be measured on outcomes by group and mitigated.
Clue to watch for Outcome rates still differ by group, or a location or similar feature ranks highly.
-
Pseudonymized means anonymous
- What candidates pick
- Share patient records after replacing names with random IDs.
- Why it looks right
- No names appear in the data.
- Why it's wrong
- Pseudonymized data is still personal data under GDPR, and quasi-identifiers can re-identify people; minimisation and stronger de-identification are needed.
Clue to watch for Detailed attributes such as birth date, sex and postal code remain.
-
On-chain is anonymous
- What candidates pick
- Publish analysis of individual wallet behavior because addresses carry no names.
- Why it looks right
- Blockchains are often described as anonymous.
- Why it's wrong
- Addresses are pseudonymous; clustering and links to exchanges or social posts can identify owners, and the public record is permanent.
Clue to watch for Analysis targets individual addresses rather than aggregates.
-
Accuracy beats explainability
- What candidates pick
- Deploy the most accurate deep model to decide benefit eligibility.
- Why it looks right
- Higher accuracy seems to serve applicants best.
- Why it's wrong
- High-stakes decisions about individuals require explanation and human review; an interpretable model or robust explanations with oversight is the responsible choice.
Clue to watch for Individuals are significantly affected and may contest the decision.
-
Observational proof of impact
- What candidates pick
- Compare users who adopted a new feature with those who did not to prove it raises spending.
- Why it looks right
- The data is already available and large.
- Why it's wrong
- Adopters differ from non-adopters in ways that drive spending; a randomized A/B test isolates the feature's causal effect.
Clue to watch for The goal is to prove a causal effect and random assignment is feasible.
-
Collect everything
- What candidates pick
- Gather every available customer attribute to maximize model accuracy.
- Why it looks right
- More features can improve prediction.
- Why it's wrong
- Data minimisation and purpose limitation require collecting only what a defined purpose needs, and extra attributes add privacy and proxy-bias risk.
Clue to watch for The scenario mentions personal data and no stated purpose for the extra fields.
Answer framework
Use these steps on applied case and ethics scenarios.
- Frame the business problem and the industry context, noting the key constraint the scenario plants, such as sensitive health data, a protected group or a public ledger.
- Choose the method and evaluation that fit the domain, such as recall for rare costly misses, randomized tests for causal claims and time-aware validation for seasonal demand.
- Check privacy: lawful purpose, data minimisation, whether data is truly anonymous or only pseudonymous, and re-identification risk from quasi-identifiers.
- Check fairness by measuring outcomes across groups, looking for proxies and naming the bias source before choosing a mitigation.
- Confirm accountability: a named owner, documentation, explainability suited to the stakes and human review for decisions affecting individuals.
Practice question
Want more like this? The EBDS practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with topic 9.3, data ethics, because its ideas decide answers across the whole module and the case studies. Build a short table of privacy concepts (anonymous, pseudonymous, minimisation, purpose limitation), bias sources and fairness metrics, and practice spotting the planted constraint in a scenario. High-priority concepts are proxy variables, pseudonymization and explainability for high-stakes decisions.
Next, work through topic 9.2 by industry. For each case, write the problem framing, the method you would choose from earlier modules, the evaluation metric and the main ethical or regulatory risk. This forces the integration the exam expects and exposes weak modules to revisit.
Topic 9.1 is commonly underestimated because Web3 seems peripheral. Learn what makes on-chain data different (public, permanent, pseudonymous) and how that changes privacy and analysis. Afterwards, take full-length timed mock exams, since this module draws on all nine, and review every miss by asking whether you ignored a constraint in the scenario.
FAQ
Why is applied data science the last EBDS module?
It integrates the whole course: case studies need every earlier technique, and ethics questions judge whether a technically sound solution is also lawful, fair and accountable.
How many questions come from this module?
The Alliance does not publish module weights. An even split of the 80 questions suggests about nine, though its ideas appear in scenarios across the paper.
Do I need to know blockchain technology in depth?
Expect concept-level questions: how on-chain data is structured and accessed, and what transparency, pseudonymity and immutability mean for analysis and privacy.
Is removing protected attributes enough to make a model fair?
No. Proxy variables can reproduce the same bias, so fairness must be measured on outcomes across groups and then mitigated.
What is the difference between anonymized and pseudonymized data?
Anonymized data cannot reasonably be linked back to individuals; pseudonymized data replaces identifiers but can be re-linked and remains personal data under GDPR.
Which fairness metric should I use?
It depends on the harm you most need to prevent. Demographic parity equalizes outcome rates, equal opportunity equalizes true positive rates, and equalized odds also equalizes false positive rates; they can conflict.
How should I prepare for the case-study questions?
Practice framing each case end to end: problem, method, metric, communication and ethical risk. Most wrong options fail on one of those steps, not on the algorithm.
Other Data & Analytics certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 9 knowledge
You have covered Web3 data, industry case studies and data ethics across privacy, fairness and accountability. Check whether you can choose answers that are technically sound and responsible in full exam scenarios.