CDSP (DSP-110) flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

SMART goal criteria

A goal framing requiring Specific metric, Measurable data, Agreed owner, Relevant decision link, and Time bound; missing elements indicate the problem is not yet defined.

Cost-of-error framing

Quantifying the business cost of false positives and false negatives separately; their ratio determines the appropriate classification threshold and acceptable error balance.

Baseline comparison rule

The current process performance must be measured and stated before modeling; a model earns adoption only by beating this status quo after accounting for its cost.

KPI ladder

A hierarchy of north-star metric, driver metrics, and guardrail metrics; each model objective is mapped to one rung so optimization does not harm protected outcomes.

Assumption log

A record of each analytical assumption with its owner and the test that would falsify it; unresolved entries become explicit risks in the final recommendation.

Decision versus exploration

A framing rule: repeatable choices call for supervised prediction, unknown structure calls for unsupervised description, and constrained allocation calls for optimization.

MNAR missingness

Missingness that depends on the unobserved value itself, so standard imputation biases results; requires sensitivity analysis or explicit modeling of the missingness mechanism.

MAR missingness

Missingness explained by observed variables, so model-based imputation using those predictors is valid and generally preferable to simple mean substitution.

Data leakage

Use of information unavailable at prediction time, such as statistics fitted before splitting or future values in features; it inflates validation scores and collapses in production.

IQR outlier fence

A rule flagging values below Q1 minus 1.5 times the interquartile range or above Q3 plus 1.5 times it; flagged points still require judgment about error versus signal.

SMOTE resampling

Synthetic minority oversampling that interpolates between minority neighbors; it alters the class prior, so predicted probabilities need recalibration after application.

Standardization versus normalization

Z-score standardization centers data at zero with unit variance and is outlier-sensitive; min-max normalization maps to a fixed range and is even more sensitive to extremes.

Chronological train-test split

Splitting time-ordered data so training precedes testing; random shuffling leaks future information into training and overstates expected production performance.

Data quality dimensions

Completeness, uniqueness, validity, consistency, timeliness, and accuracy; each is checked against explicit rules before analysis begins.

Bias-variance tradeoff

High bias means both training and validation error are high; high variance means low training error with a large validation gap, remedied by regularization or more data.

Precision-recall versus ROC

Precision-recall curves and PR-AUC are preferred when positives are rare; ROC-AUC stays optimistic under severe imbalance because it includes abundant true negatives.

L1 versus L2 regularization

L1 lasso shrinks some coefficients exactly to zero and selects features; L2 ridge shrinks smoothly and is preferred when predictors are correlated.

Stratified k-fold

Cross-validation that preserves the class proportion in every fold; necessary for classification, especially with imbalance, to keep fold metrics comparable.

RMSE versus MAE

RMSE is in target units and penalizes large errors heavily; MAE is the average absolute error and is more robust when outliers are present.

Permutation importance

A post-hoc measure that shuffles a feature and records the performance drop; it explains model reliance, not causal effect, and is unreliable with correlated features.

Overfitting signals

Validation performance degrading while training performance improves, or early learning-curve plateau, indicates the model captures noise rather than generalizable structure.

Regression residual assumptions

Linearity, independent errors, constant variance, and normal residuals; violations suggest transformation, robust standard errors, or a different model family.

Confidence interval meaning

A 95 percent interval means the procedure captures the true parameter in 95 percent of repeated samples, not that this interval holds it with 95 percent probability.

p-value definition

The probability of observing data at least as extreme as collected, assuming the null hypothesis is true; it is not the probability the null is true.

Statistical versus practical significance

Large samples make trivial effects statistically significant; always report effect size against a business-relevant threshold alongside the p-value.

Simpson's paradox

An association present in aggregated data can reverse within every subgroup; check segment-level direction before reporting an overall effect.

Multiple comparisons correction

Testing many hypotheses inflates false positives; Bonferroni divides alpha by the number of tests, while Benjamini-Hochberg controls the false discovery rate.

Extrapolation risk

Model validity is bounded by the predictor range and population seen in training; predictions outside that support region are unvalidated and should be flagged.

Experiment power failure

An inconclusive result from insufficient sample size differs from a genuine null; distinguish them before concluding the intervention has no effect.

Answer-first communication

A structure leading with the decision-relevant conclusion, then supporting evidence, then method detail, so stakeholders encounter the recommendation immediately.

Audience mapping

Tailoring content by role: executives need decision, impact, and risk; analysts need method and assumptions; operators need the action and its trigger conditions.

Chart selection rule

Bars for category comparison, lines for trends over time, scatter for two-variable relationships, histograms for distributions; pie charts only for few-part wholes.