Hard Difficulty Questions

Consistency

Why it fails: Consistency concerns conflicting values across sources or formats, not repeated identifiers within a dataset. It is tempting because duplicates look like conflicting records, but it would be correct when the same attribute disagrees between systems; uniqueness is the dimension violated by duplicate customer IDs.

Completeness

Why it fails: Completeness concerns missing values or records, not repeated customer IDs. It is tempting because duplicates feel like missing distinctness, but it would be correct when required fields or rows are absent; uniqueness is the dimension breached by duplicate identifiers.

Accuracy

Why it fails: Duplicate customer IDs breach uniqueness, not accuracy; accuracy concerns whether a value correctly describes the real-world entity. It tempts because duplicate records often contain stale or conflicting attributes, so accuracy problems frequently accompany them, but the dimension tested here is the one-to-one mapping of each record to a distinct customer.

Visual hierarchy

Why it fails: Visual hierarchy concerns arranging elements so the most important metrics draw attention first. Decorative graphics and background images instead clutter the canvas with non-data ink, lowering the data-ink ratio. Visual hierarchy would be the answer if key KPIs were buried beneath less important widgets.

Appropriate precision

Why it fails: Decorative graphics and background images violate the principle of maximising the data-ink ratio, not precision. Precision concerns rounding and significant figures in displayed values. It would be the correct answer if the dashboard showed figures such as 47.38291% where two decimal places suffice.

Consistent color coding

Why it fails: Consistent colour coding governs using the same hue for the same category across charts. Decorative graphics and background images instead reduce the data-to-ink ratio by adding non-data pixels. Colour coding would be the answer if the same metric appeared in different colours on separate dashboard tiles.

Lasso regression (L1)

Why it fails: Lasso applies L1 regularisation to shrink coefficients, not to select polynomial degree; cross-validation is the technique for choosing degree. Lasso suits high-dimensional linear models with irrelevant predictors, a different problem from overfitting control in polynomial fitting.

Principal component analysis (PCA)

Why it fails: PCA reduces dimensionality by projecting onto principal components; it does not evaluate polynomial degree or detect overfitting. PCA suits correlated feature sets before modelling, whereas degree selection requires cross-validation comparing validation error across candidate degrees.

Stepwise selection

Why it fails: Stepwise selection adds or removes predictors by an information criterion; it does not tune polynomial degree, which is set by cross-validation error across candidate degrees. It is tempting because stepwise selection genuinely suits choosing among many candidate input variables in multiple regression.

Reject the entire dataset

Why it fails: Rejecting the dataset is drastic; the issues are fixable.

Ignore warnings and continue

Why it fails: Ignoring warnings may lead to further data quality issues.

Correct the date string in the source

Why it fails: Fixing a single row does not address the systematic type mismatch.

Impute using regression.

Why it fails: Regression imputation is premature without understanding the data quality issue.

Replace negative age with the mean age.

Why it fails: Imputation should follow root cause analysis to avoid masking data quality problems.

Remove all records with negative age.

Why it fails: Removing data without investigation can discard valuable information and hides underlying issues.

Keep both as is and use a transformation layer.

Why it fails: Maintaining two formats complicates queries and analysis.

Store all amounts as strings to preserve precision.

Why it fails: Strings lose numeric functionality for calculations.

Convert System B amounts to cents by multiplying by 100.

Why it fails: This also standardizes but results in larger integer values; while acceptable, option A is more intuitive for financial analysis.

Download the file to a staging server, then bulk insert into warehouse

Why it fails: Download adds an extra step and uses storage.

Have the vendor push data via API instead of SFTP

Why it fails: Changing vendor processes is not always feasible.

Split the file into smaller chunks and load concurrently

Why it fails: Splitting still requires download and adds complexity.

Emphasize the statistical significance of the finding and ignore previous reports

Why it fails: Ignoring previous reports undermines credibility and does not address the audience's concerns.

Present the data without any explanation and let them draw conclusions

Why it fails: Without explanation, the audience may dismiss the finding due to the discrepancy.

Remove the demographic detail to avoid controversy

Why it fails: Removing details hides important insights and could be seen as dishonest.

Switch to a line chart with each server as a separate line

Why it fails: Plotting 200 separate lines still overlaps heavily and colours remain indistinguishable, so over-80% servers and spikes stay hidden. Line charts suit a small number of series tracked over time, not 200 servers at five-minute granularity.

Add a trend line to each server's data and remove the individual points

Why it fails: A trend line per server removes the individual readings, so unusual spikes cannot be detected, and 200 trend lines remain cluttered. Trend lines suit showing long-term direction in a few series, not spike detection across many servers.

Increase the size of the data points to make them more visible

Why it fails: Larger markers increase overlap and clutter.

Data dictionary

Why it fails: A data dictionary documents technical metadata — field names, data types, formats and lineage — rather than mandating classification tiers or handling rules. It would be the correct artefact when cataloguing schema definitions, whereas policy-level classification belongs in a data classification policy or standard.

Data quality report

Why it fails: A data quality report measures accuracy, completeness and consistency of datasets; it defines no classification tiers or handling procedures, so it cannot govern PII. It tempts because quality reporting supports governance oversight, and it would be the right artefact when the requirement is monitoring data fitness rather than prescribing handling rules.

Data flow diagram

Why it fails: A data flow diagram maps where data moves between systems and processes; it records no classification tiers or handling rules, so it cannot serve as the governing policy. It is tempting because flow diagrams are genuinely useful for tracing PII propagation during a privacy impact assessment, but the classification scheme itself belongs in a data classification policy.

Switch to a bar chart sorted by sales

Why it fails: A sorted bar chart abandons the geographic layout entirely, so spatial patterns across states are lost rather than corrected. It tempts because bars remove area bias and rank values clearly, yet the task is a map-based per-capita comparison, which a cartogram preserves.

Use a bubble chart with size proportional to sales

Why it fails: Bubble size encodes absolute sales, so Delaware's high per-capita figure still inflates its marker; the geographic-area distortion remains. It tempts because bubbles decouple value from land area, but the stem requires per-capita comparison, which a cartogram or proportional symbol of population handles.

Apply a log scale to the color gradient

Why it fails: A log scale compresses the colour gradient's numeric range but leaves each state's polygon area unchanged, so Delaware still dominates visually. It tempts because log scales do tame skewed values, yet the distortion here is geographic area, not magnitude.

A parallel coordinates plot displaying all variables used in clustering.

Why it fails: Parallel coordinates are complex and difficult for non-technical viewers to interpret.

A heatmap of the distance matrix between cluster centroids.

Why it fails: Heatmaps of distances are abstract and not suitable for explaining cluster profiles.

A scatter plot with two principal components showing cluster boundaries.

Why it fails: PCA and scatter plots are technical and may confuse non-technical audiences.

Cyclical

Why it fails: Cyclical patterns fluctuate over multi-year, non-fixed periods tied to economic or business cycles, not regular weekly or yearly intervals. The term tempts because it also describes repetition over time, but its lack of a fixed period distinguishes it from the seasonal pattern the stem describes.

Irregular

Why it fails: Irregular components are random, unsystematic residuals left after trend, seasonal and cyclical components are removed; they show no regular recurrence. The label tempts because irregular events do recur, but without a fixed interval, so they cannot match the weekly or yearly periodicity the stem specifies.

Trend

Why it fails: A trend is the long-term direction of the series — sustained upward or downward movement — rather than a pattern repeating at fixed intervals. Trend tempts because it is a core decomposition component, yet it captures direction over time, not the regular weekly or yearly cycle described.

Increase polynomial features

Why it fails: This would likely increase variance further.

Use a smaller training set

Why it fails: Smaller training set typically increases variance.

Remove correlated features

Why it fails: This may reduce variance but can increase bias; not the most direct method.

Data can be categorized into groups

Why it fails: Categorization is a property of nominal/ordinal data, not specific to ratio.

Allows negative values

Why it fails: True zero implies no negative values (e.g., weight cannot be negative).

No violation

Why it fails: Storing Product1, Product1Qty, Product2 and Product2Qty as repeating column groups places multiple values in a single attribute position, which breaches first normal form. Claiming no violation is tempting when OrderID appears to be a clean primary key, but key uniqueness does not resolve the multivalued columns that 1NF forbids.

Third normal form (3NF)

Why it fails: The repeating Product1/Product2 groups breach first normal form by storing multiple values per attribute, so the design never reaches the stage 3NF addresses. 3NF is tempting because it targets transitive dependencies between non-key columns, which would be the correct diagnosis if each product occupied its own atomic row and CustomerName depended on a non-key attribute.

Second normal form (2NF)

Why it fails: 2NF is about partial dependencies, but 1NF is violated first.

Data is processed in micro-batches

Why it fails: Micro-batch is a style, but not a defining characteristic of streaming.

Data is stored permanently before processing

Why it fails: Streaming data is typically processed on arrival.

Increase the request rate

Why it fails: Exceeding 100 requests per minute triggers HTTP 429 responses and possible throttling or blocking, so the data is not retrieved faster. It is tempting because raising throughput is the obvious lever when latency, not a server-enforced quota, is the limiting factor.

Use multiple API keys

Why it fails: Rotating keys circumvents the provider's quota and typically breaches the API terms, risking suspension of every key. It is tempting because separate keys appear to grant independent allowances, which works only where the provider legitimately issues per-application quotas.

Download a precompiled dataset

Why it fails: A precompiled dataset abandons the API entirely, so it cannot satisfy a requirement to collect the specified historical records through that API. It is tempting because bulk downloads avoid rate limits completely, which suits exploratory analysis where any equivalent dataset suffices.

Remove poorly calibrated predictions by discarding all patients with predicted risk between 0.3 and 0.7.

Why it fails: Discarding patients with predicted risk between 0.3 and 0.7 deletes the very cases needing recalibration and biases the test set, leaving the underlying probability estimates unchanged. It is tempting because restricting to confident predictions appears to remove miscalibrated outputs, but Platt scaling fits a logistic transform on held-out scores instead.

Ignore calibration because AUC is the only metric that matters for readmission risk models.

Why it fails: Calibration matters clinically because risk thresholds drive intervention decisions; a 70% prediction that is truly 40% misleads care teams. AUC only measures ranking, so ignoring calibration leaves the stated defect unresolved. It is tempting when discrimination is the sole reported metric, but probability reliability is the requirement here.

Switch to a random forest model, which inherently produces better-calibrated probabilities.

Why it fails: Random forests produce votes averaged across trees, which are typically pushed toward 0 and 1 and are therefore often poorly calibrated, not inherently reliable. It is tempting because ensembles are assumed to output trustworthy probabilities, but Platt scaling directly fits a sigmoid to the existing logistic regression scores without changing the algorithm.

The data type of order_id is incorrect.

Why it fails: No type conversion error.

The source and target schemas are mismatched.

Why it fails: No schema mismatch error is shown.

The order_id field contains null values.

Why it fails: No null violation error.