Dell Data Science Associate flashcards

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

CRISP-DM

A six-phase, iterative data science methodology: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. It emphasizes returning to earlier phases as findings change requirements.

Bias-variance trade-off

Expected prediction error decomposes into squared bias, variance, and irreducible noise. Complex models lower bias but raise variance; simpler models do the reverse. Balance is tuned via validation performance.

Precision

The fraction of predicted positives that are truly positive: TP divided by TP plus FP. High precision means few false alarms, valuable when false positives carry high cost.

Recall (sensitivity)

The fraction of actual positives correctly identified: TP divided by TP plus FN. High recall means few missed cases, valuable when false negatives are costly, such as disease screening.

F1 score

The harmonic mean of precision and recall: 2PR/(P+R). It penalizes lopsided models and is useful for imbalanced classification where accuracy alone misleads.

ROC-AUC

The probability that a randomly chosen positive case receives a higher predicted score than a randomly chosen negative case. It is threshold-independent and equals 0.5 for random guessing.

Data leakage

Use of information during training that will not be available at prediction time, such as target-derived features or future records. It inflates validation scores and causes silent production failure.

Nominal vs ordinal data

Nominal categories have no order, like color or region. Ordinal categories have a meaningful order but unequal or undefined spacing, like satisfaction ratings from low to high.

MCAR, MAR, MNAR

Missing-data mechanisms: missing completely at random (independent of all values), missing at random (explained by observed variables), and missing not at random (depends on the missing value itself).

One-hot encoding

Converting a categorical variable into binary indicator columns, one per category. For linear models, drop one column (k-1) to avoid perfect multicollinearity, the dummy variable trap.

Standardization (z-score)

Rescaling a feature as (x minus mean) divided by standard deviation, giving mean zero and unit variance. Required for distance-based and regularized models; unnecessary for tree ensembles.

Min-max scaling

Rescaling a feature to a fixed range, typically zero to one, using (x minus min) divided by (max minus min). Sensitive to outliers because extremes set the boundaries.

IQR outlier rule

Flag values below Q1 minus 1.5 times the interquartile range or above Q3 plus 1.5 times it. The rule is distribution-free and robust compared with z-score thresholds.

Third normal form (3NF)

A relational table design where values are atomic, every non-key attribute depends on the whole key, and no non-key attribute depends transitively on another non-key attribute.

WHERE versus HAVING

WHERE filters individual rows before grouping and aggregation. HAVING filters aggregated groups after GROUP BY, so it can reference aggregate functions like SUM or COUNT.

Idempotent pipeline

A data pipeline that produces the same result whether run once or many times, typically by overwriting partitions or using merge logic instead of blind appends.

Multiple imputation

Replacing missing values several times with plausible draws, analyzing each completed dataset, then pooling estimates and standard errors. Preferred over single imputation under MAR.

Ordinary least squares

A linear regression fitting method that chooses coefficients minimizing the sum of squared residuals. It yields closed-form solutions and unbiased estimates under the standard assumptions.

Adjusted R-squared

R-squared penalized for the number of predictors, so it can decrease when useless variables are added. Preferred over plain R-squared for comparing models with different feature counts.

Logistic regression

A classifier modeling the log-odds of a binary outcome as a linear combination of features. Exponentiating a coefficient gives an odds ratio for a one-unit change.

Lasso (L1) regularization

Adds the sum of absolute coefficient values to the loss, shrinking some coefficients exactly to zero. It performs automatic feature selection and suits sparse high-dimensional problems.

Ridge (L2) regularization

Adds the sum of squared coefficients to the loss, shrinking coefficients toward zero without eliminating them. It stabilizes estimates when predictors are highly correlated.

k-fold cross-validation

Partitioning data into k folds, training on k-1 and validating on the remaining fold, rotating until each fold serves once, then averaging scores. Reduces reliance on one split.

Random forest

An ensemble of decision trees trained on bootstrap samples with random feature subsets at each split. Averaging decorrelated trees reduces variance and overfitting relative to a single tree.

Gradient boosting

An ensemble built sequentially, where each new tree fits the residual errors of the current model. It typically achieves low bias but needs careful learning rate and depth tuning.

k-means clustering

An unsupervised algorithm partitioning points into k clusters by minimizing within-cluster squared distances. It requires k in advance and assumes roughly spherical, similarly sized clusters.

p-value

The probability of observing data at least as extreme as the sample, assuming the null hypothesis is true. It is not the probability that the null is true.

Confidence interval

A range of plausible values for a population parameter, constructed so that repeated sampling would capture the true value at a stated rate, commonly 95 percent.

Chart type selection

Bars compare categories, lines show trends over time, scatter plots reveal relationships between two variables, and histograms show distributions. Pie charts suit few parts of one whole.

Statistical vs practical significance

A small p-value shows an effect is unlikely due to chance, not that it is large or valuable. Report effect size and confidence intervals to judge business relevance.

A/B test randomization

Randomly assigning units to control and treatment so groups differ only by chance. It supports causal claims about the intervention, unlike pre/post observational comparisons.

KPI definition

A metric specified by numerator, denominator, time window, and exclusions, so different teams compute it identically and results remain comparable across reporting periods.

Minimum detectable effect

The smallest true effect a test is powered to detect at a given sample size, significance level, and power. It is fixed during test design and drives required sample size.