CompTIA Data+ cheat sheet

Structural tiers: structured data fits a fixed tabular schema (RDBMS, schema-on-write); semi-structured data uses keys or tags such as JSON/XML (schema-on-read); unstructured data has no predefined model (text, images, audio, video).

Normalization rules: 1NF requires atomic column values and no repeating groups; 2NF removes partial dependencies on a composite key; 3NF removes transitive dependencies between non-key columns.

Denormalization intentionally reintroduces redundancy (flattened or widened tables) to cut join count and speed read-heavy BI queries at the cost of write integrity.

Dimensional model: a fact table stores numeric measures plus foreign keys, while dimension tables store the descriptive attributes used for filtering, grouping, and labeling.

Star schema keeps dimensions denormalized (fewer joins, faster queries); snowflake schema normalizes dimensions into multiple related tables (less redundancy, more joins).

Measurement scales: nominal = unordered labels; ordinal = ordered ranks with unequal intervals; interval = equal intervals but no true zero; ratio = equal intervals plus a true zero, so ratios are valid.

Cardinality rules: 1:1 and 1:many are modeled with a foreign key on the many side; many:many requires an associative (bridge) table holding both keys.

Storage targets: a data lake holds raw data in native format at scale; a data warehouse holds modeled, cleansed data for reporting; a data mart is a warehouse subset serving one business unit.

Common data types: integer and decimal/numeric are exact; float/double are approximate binary representations and should not be used for currency; char is fixed length, varchar variable length.

CRISP-DM sequence: business understanding → data understanding → data preparation → modeling → evaluation → deployment; the cycle is iterative and commonly loops back after evaluation.

Task-to-algorithm rule: predict a discrete class = classification; predict a continuous number = regression; find natural groups without labels = clustering; find co-occurring items = association rules.

k-means procedure: choose k, seed k centroids, assign each point to the nearest centroid, recompute centroids as the mean of assigned points, repeat until assignments stop changing; requires scaled features.

Decision tree splits are chosen by impurity reduction, using Gini impurity or entropy (information gain); unrestricted depth overfits, so pre-pruning (max depth, min samples) or post-pruning is applied.

Classification metrics: accuracy = (TP+TN)/total; precision = TP/(TP+FP); recall = TP/(TP+FN); F1 = 2·(precision·recall)/(precision+recall); specificity = TN/(TN+FP).

Overfitting signature: training error stays low while validation error rises. Remedies are more or better data, feature reduction, a simpler model, regularization, and cross-validation.

Validation splits: a holdout of 70/30 or 80/20 is typical; k-fold cross-validation (k = 5 or 10) rotates the test fold so every record is used for training and testing once.

Association rules: support(X) = transactions containing X ÷ total transactions; confidence(X→Y) = support(X∪Y)/support(X); lift = confidence ÷ support(Y), where lift > 1 means positive association.

Outlier decision rules: flag values beyond Q1 − 1.5·IQR or Q3 + 1.5·IQR; flag z-scores with |z| > 3 - these rules assume the underlying distribution is roughly normal.

Analytics types: descriptive answers 'what happened'; diagnostic answers 'why'; predictive forecasts 'what will happen'; prescriptive recommends 'what action to take'.

Central tendency: the mean is the arithmetic average and is distorted by outliers or skew; the median is the middle ordered value and is robust; the mode is the most frequent value and is the only valid measure for nominal data.

Spread measures: range = max − min; variance is the mean squared deviation; standard deviation is its square root in original units; coefficient of variation = standard deviation ÷ mean, used to compare variability across different units.

Correlation coefficient r ranges from −1 to +1, where 0 indicates no linear relationship; r measures linear association only and never establishes causation.

Hypothesis testing: state H0 and H1, choose alpha (typically 0.05), compute the test statistic and p-value; if p < alpha, reject H0. Compare two means with a t-test, three or more means with ANOVA, and two categorical variables with chi-square.

Type I error = false positive (rejecting a true null), probability equal to alpha; Type II error = false negative (failing to reject a false null), probability beta; power = 1 − beta.

Simple linear regression: y = mx + b, where m is the slope and b the intercept; the residual is actual minus predicted, and R² is the proportion of variance in y explained by the model.

Join semantics: inner join returns only matching keys in both tables; left join returns all left rows plus matches (nulls where unmatched); full outer returns all rows from both; cross join returns the Cartesian product.

SQL filtering order: WHERE filters individual rows before aggregation, HAVING filters aggregated groups after GROUP BY, and both must precede ORDER BY in the logical evaluation sequence.

Chart selection rule: compare categories = bar; trend over time = line; distribution of one numeric field = histogram or box plot; relationship between two numerics = scatter; composition = stacked bar or treemap; location = map.

Pie chart constraints: slices must sum to 100% of a single whole, categories should be limited to roughly five or fewer, and ordering by size makes small differences readable.

Baseline rule: bar charts must start at zero because bar length encodes magnitude; line charts may use a non-zero baseline when the axis is clearly labeled, since they encode change rather than magnitude.

Box plot anatomy: box spans Q1 to Q3, the line inside is the median, whiskers extend to 1.5·IQR, and points beyond the whiskers are plotted as outliers.

Histogram binning: bins must be equal width and adjacent; bin count controls apparent shape - too few hides structure, too many shows noise; bins differ from bar-chart categories because they represent continuous ranges.

Color scale selection: sequential for ordered magnitude, diverging for deviation around a meaningful midpoint, and categorical (qualitative) for nominal groups with no inherent order.

Accessibility rule: never rely on color alone - pair hue with labels, patterns, or direct annotations, and avoid red/green contrasts as the only differentiator.

Chart junk reduction: maximize the data-ink ratio by removing gridlines, 3D effects, borders, and decorative images, and by labeling axes with units and a clear title.

Waterfall charts show the cumulative effect of sequential positive and negative contributions, making them the standard view for variance or budget-to-actual bridge analysis.

Governance roles: the data owner is accountable for a data domain and approves access; the data steward enforces standards and resolves quality issues day to day; the data custodian handles technical storage, backup, and security.

Data quality dimensions: accuracy, completeness, consistency, timeliness, validity, and uniqueness - each is measured against a defined threshold and tracked over time.

Master data management creates one authoritative, de-duplicated record for core entities such as customer, product, and supplier, and synchronizes it across systems.

Data lineage records the origin, movement, and transformation of data from source to report; provenance records where a record came from - both support impact analysis and audit.

Sensitive categories: PII identifies or can re-identify a person (name, SSN, email); PHI is health information tied to an individual; both require controls over access, transmission, and retention.

Privacy rights under GDPR and CCPA include data subject access requests, correction, deletion (right to be forgotten), portability, and limits on retention and sale of personal data.

Access control models: RBAC grants permissions by role; ABAC grants by attributes such as department or clearance; both implement least privilege and need-to-know.

De-identification techniques: masking replaces values with fictitious substitutes; tokenization swaps a value for a non-reversible reference token; pseudonymization keeps re-identification possible via a key; anonymization is irreversible.

Regulation mapping: HIPAA covers protected health information in the US; PCI DSS covers cardholder data; SOX covers financial reporting controls; FERPA covers education records.