Dell Data Science Associate cheat sheet

CRISP-DM phases in order: business understanding, data understanding, data preparation, modeling, evaluation, deployment; the process is iterative, not linear.

Bias-variance decomposition: expected squared error = bias^2 + variance + irreducible noise; high bias underfits, high variance overfits.

Supervised learning requires labeled targets; unsupervised learning (clustering, dimensionality reduction) does not; reinforcement learning learns from reward signals via exploration.

Confusion matrix derived metrics: accuracy = (TP+TN)/(TP+TN+FP+FN); precision = TP/(TP+FP); recall = TP/(TP+FN); F1 = 2precisionrecall/(precision+recall).

Precision matters when false positives are costly (spam filters); recall matters when false negatives are costly (disease screening).

ROC-AUC equals the probability a random positive is scored above a random negative; it is threshold-independent, while precision-recall curves suit heavily imbalanced data.

Correlation measures linear association only; r = 0 does not imply independence. Association is not causation without randomization or a defensible causal design.

Types of data: nominal (unordered categories), ordinal (ordered categories), interval (equal differences, no true zero), ratio (true zero permits ratios).

Data leakage occurs when information unavailable at prediction time (e.g., target-derived features, future data) enters training, inflating validation scores.

Missing data mechanisms: MCAR (missing completely at random), MAR (missingness explained by observed variables), MNAR (depends on the unobserved value itself); listwise deletion is unbiased only under MCAR.

Imputation choices: mean/median preserves the mean but shrinks variance; add a missing-indicator column to retain missingness signal; multiple imputation is preferred for MAR.

Scaling rules: standardize (z = (x - mean)/sd) for distance- and gradient-based models; min-max scale to [0,1] for bounded inputs; tree ensembles need no scaling.

One-hot encode nominal variables with k-1 columns for linear models to avoid perfect multicollinearity (dummy variable trap); high-cardinality categories favor target or frequency encoding.

Outlier detection: IQR rule flags points below Q1 - 1.5IQR or above Q3 + 1.5IQR; z-score thresholds assume approximate normality.

Normalization forms: 1NF atomic values, 2NF no partial dependency on a composite key, 3NF no transitive dependency on non-key attributes.

SQL execution order: FROM/JOIN, WHERE, GROUP BY, HAVING, SELECT, ORDER BY, LIMIT; WHERE filters rows before aggregation, HAVING filters after.

Joins: INNER returns matches only; LEFT keeps all left rows with NULLs for unmatched; FULL keeps all rows; a join on non-unique keys multiplies rows.

Idempotent pipelines can be rerun without duplicating results; partition pruning and predicate pushdown reduce scan volume in columnar stores.

Linear regression assumptions: linearity, independent errors, homoscedasticity, normally distributed residuals, no severe multicollinearity; coefficients are interpreted as change in y per unit change in x holding others constant.

Ordinary least squares minimizes the sum of squared residuals; R-squared is the explained variance fraction and never decreases when predictors are added, so use adjusted R-squared or cross-validation.

Logistic regression models log-odds as a linear function: log(p/(1-p)) = b0 + b1*x1 + ...; coefficients exponentiate to odds ratios.

Bias-variance trade-off via regularization: L1 (lasso) drives coefficients to zero and selects features; L2 (ridge) shrinks coefficients; elastic net combines both.

k-fold cross-validation: split into k folds, train on k-1, validate on the held-out fold, average scores; stratified folds preserve class proportions for imbalanced targets.

Decision trees split on impurity reduction (Gini or entropy); random forests average many decorrelated trees to cut variance; gradient boosting fits sequential trees to residual errors.

k-means minimizes within-cluster sum of squares, requires k in advance, and assumes roughly spherical clusters; choose k via elbow or silhouette score.

Hypothesis testing: p-value is the probability of data at least as extreme given the null is true, not the probability the null is true; alpha is chosen before testing.

Class imbalance remedies: resampling, class weights, threshold tuning, and metrics such as precision-recall AUC or lift rather than raw accuracy.

Audience-first structure: lead with the decision or recommendation, then supporting evidence, then method detail; executives need the 'so what' before the 'how'.

Chart selection: bar charts compare categories, line charts show trends over time, scatter plots show relationships, histograms show distributions; pie charts are limited to a few parts of one whole.

Truncated y-axes exaggerate differences; dual axes invite spurious correlation claims; label units, sources, and time windows on every visual.

KPI definitions must be explicit (numerator, denominator, period, exclusions) so metrics are reproducible across teams.

Correlation-versus-causation framing: observational lifts are associations; only controlled experiments (A/B tests) support causal claims about interventions.

A/B test design: predefine the primary metric and minimum detectable effect, randomize assignment, and compute required sample size before launch to avoid peeking bias.

Statistical significance differs from practical significance; report effect size and confidence intervals alongside p-values to convey business impact.

Data storytelling arc: context and business question, insight with evidence, recommendation with expected impact, and risks or caveats.

Dashboards should map to decisions: each panel answers a specific question, with freshness timestamps and clear ownership.

CRISP-DM phases in order: business understanding, data understanding, data preparation, modeling, evaluation, deployment; the process is iterative, not linear.

Bias-variance decomposition: expected squared error = bias^2 + variance + irreducible noise; high bias underfits, high variance overfits.

Supervised learning requires labeled targets; unsupervised learning (clustering, dimensionality reduction) does not; reinforcement learning learns from reward signals via exploration.

Confusion matrix derived metrics: accuracy = (TP+TN)/(TP+TN+FP+FN); precision = TP/(TP+FP); recall = TP/(TP+FN); F1 = 2precisionrecall/(precision+recall).

Precision matters when false positives are costly (spam filters); recall matters when false negatives are costly (disease screening).

ROC-AUC equals the probability a random positive is scored above a random negative; it is threshold-independent, while precision-recall curves suit heavily imbalanced data.

Correlation measures linear association only; r = 0 does not imply independence. Association is not causation without randomization or a defensible causal design.

Types of data: nominal (unordered categories), ordinal (ordered categories), interval (equal differences, no true zero), ratio (true zero permits ratios).

Data leakage occurs when information unavailable at prediction time (e.g., target-derived features, future data) enters training, inflating validation scores.

Missing data mechanisms: MCAR (missing completely at random), MAR (missingness explained by observed variables), MNAR (depends on the unobserved value itself); listwise deletion is unbiased only under MCAR.

Imputation choices: mean/median preserves the mean but shrinks variance; add a missing-indicator column to retain missingness signal; multiple imputation is preferred for MAR.

Scaling rules: standardize (z = (x - mean)/sd) for distance- and gradient-based models; min-max scale to [0,1] for bounded inputs; tree ensembles need no scaling.

One-hot encode nominal variables with k-1 columns for linear models to avoid perfect multicollinearity (dummy variable trap); high-cardinality categories favor target or frequency encoding.

Outlier detection: IQR rule flags points below Q1 - 1.5IQR or above Q3 + 1.5IQR; z-score thresholds assume approximate normality.

Normalization forms: 1NF atomic values, 2NF no partial dependency on a composite key, 3NF no transitive dependency on non-key attributes.

SQL execution order: FROM/JOIN, WHERE, GROUP BY, HAVING, SELECT, ORDER BY, LIMIT; WHERE filters rows before aggregation, HAVING filters after.

Joins: INNER returns matches only; LEFT keeps all left rows with NULLs for unmatched; FULL keeps all rows; a join on non-unique keys multiplies rows.

Idempotent pipelines can be rerun without duplicating results; partition pruning and predicate pushdown reduce scan volume in columnar stores.

Linear regression assumptions: linearity, independent errors, homoscedasticity, normally distributed residuals, no severe multicollinearity; coefficients are interpreted as change in y per unit change in x holding others constant.

Ordinary least squares minimizes the sum of squared residuals; R-squared is the explained variance fraction and never decreases when predictors are added, so use adjusted R-squared or cross-validation.

Logistic regression models log-odds as a linear function: log(p/(1-p)) = b0 + b1*x1 + ...; coefficients exponentiate to odds ratios.

Bias-variance trade-off via regularization: L1 (lasso) drives coefficients to zero and selects features; L2 (ridge) shrinks coefficients; elastic net combines both.

k-fold cross-validation: split into k folds, train on k-1, validate on the held-out fold, average scores; stratified folds preserve class proportions for imbalanced targets.

Decision trees split on impurity reduction (Gini or entropy); random forests average many decorrelated trees to cut variance; gradient boosting fits sequential trees to residual errors.

k-means minimizes within-cluster sum of squares, requires k in advance, and assumes roughly spherical clusters; choose k via elbow or silhouette score.

Hypothesis testing: p-value is the probability of data at least as extreme given the null is true, not the probability the null is true; alpha is chosen before testing.

Class imbalance remedies: resampling, class weights, threshold tuning, and metrics such as precision-recall AUC or lift rather than raw accuracy.

Audience-first structure: lead with the decision or recommendation, then supporting evidence, then method detail; executives need the 'so what' before the 'how'.

Chart selection: bar charts compare categories, line charts show trends over time, scatter plots show relationships, histograms show distributions; pie charts are limited to a few parts of one whole.

Truncated y-axes exaggerate differences; dual axes invite spurious correlation claims; label units, sources, and time windows on every visual.

KPI definitions must be explicit (numerator, denominator, period, exclusions) so metrics are reproducible across teams.

Correlation-versus-causation framing: observational lifts are associations; only controlled experiments (A/B tests) support causal claims about interventions.

A/B test design: predefine the primary metric and minimum detectable effect, randomize assignment, and compute required sample size before launch to avoid peeking bias.

Statistical significance differs from practical significance; report effect size and confidence intervals alongside p-values to convey business impact.

Data storytelling arc: context and business question, insight with evidence, recommendation with expected impact, and risks or caveats.

Dashboards should map to decisions: each panel answers a specific question, with freshness timestamps and clear ownership.