CCNA Data Analysis Questions

Logistic regression models the probability of a binary outcome by applying the logistic (sigmoid) function to a linear combination of predictors, bounding output between 0 and 1. This directly suits the churn yes/no target, unlike ordinary linear regression.

This sequence removes formatting characters that block conversion, explicitly handles the 'N/A' entries rather than letting them fail silently, keeps negative amounts so refunds reduce revenue correctly, and performs a deliberate cast to a numeric type. It also leaves a clear audit trail of which rows were invalid, which matters when the summed total feeds a financial report.

Seasonality describes a repeating pattern occurring at fixed intervals, such as weekly or yearly cycles, distinct from trend and irregular components. This matches the stem's definition of a pattern repeating at regular intervals within time series decomposition.

R-squared is the coefficient of determination: 0.81 means 81% of the variance in the dependent variable (sales) is accounted for by the independent variable (advertising spend). The remaining 19% is unexplained by the model, so this interpretation matches the statistic precisely.

Case normalization involves converting all text to a consistent case, such as lowercase or uppercase. This resolves inconsistencies like 'Electronics', 'electronics', and 'ELECTRONICS' by making them identical. It is a simple and effective data cleaning step for standardizing categorical text values. After case normalization, the entries would all become 'electronics', for example.

Seasonality captures repeating fluctuations within a fixed period shorter than a year, so the annual December peaks and February troughs satisfy that constraint precisely. Unlike trend, which shows long-term direction, or cyclical variation, which lacks a fixed period, seasonality is defined by calendar-linked regularity.

Capping at the 99th percentile winsorises extreme values to a boundary, retaining the observation while limiting its leverage on the regression coefficients. This preserves sample size and satisfies the need to reduce outlier influence without discarding data.

K-means computes Euclidean distance, so unstandardised continuous variables with larger ranges dominate. Standardising spending and frequency equalises their influence, while one-hot encoding converts region and gender into numeric binary columns, since k-means cannot process categorical labels directly.

Logistic regression predicts a binary outcome, churn versus no churn, satisfying the categorical target requirement. It handles mixed predictors: region and plan type via dummy encoding, usage and tenure as continuous terms. Linear regression would model a continuous response, which churn is not, so logistic regression fits this classification scenario.

Log transformation compresses the long right tail by rescaling values on a multiplicative scale, pulling extreme high observations closer to the bulk. This reduces positive skewness, making the target distribution more symmetric and better suited to techniques assuming normality.

Hierarchical clustering with Gower distance computes pairwise dissimilarity across numeric and categorical attributes simultaneously, so mixed-type records can be segmented without arbitrary encoding. It satisfies the stem's mixed data constraint, unlike k-means, which relies on Euclidean distance and requires numeric, scaled input.

The median is the middle value when data is ordered, making it appropriate for ordinal data because it does not assume equal intervals. For satisfaction ratings, the median represents the central rating without implying that the difference between ratings is consistent. It is robust to outliers and skewed distributions, providing a more accurate summary of the typical response.

L2 regularization (Ridge) adds a penalty proportional to the squared magnitude of coefficients, shrinking them towards zero and reducing variance without eliminating features. This directly counters the overfitting causing the train-validation performance gap, satisfying the scenario's need to improve generalisation on unseen house price data.

One-hot encoding creates binary columns for each category, avoiding implying any ordinal relationship. For regression, it is appropriate for nominal categorical variables like Education Level, which has no inherent order. It allows the model to estimate a separate coefficient for each category, capturing differences without assuming a linear progression.

Recall, also called sensitivity or true positive rate, measures the proportion of actual defaulters that the model correctly identifies. With only 2% defaults, a model predicting 'no default' for all cases would achieve 98% accuracy but zero recall. Therefore, recall is the appropriate metric to assess the model's ability to catch actual defaulters, directly addressing the analyst's concern.

Seasonality describes repeating, calendar-linked fluctuations within a fixed period, such as December peaks and January troughs each year. The consistent annual recurrence in the monthly sales series identifies this component rather than trend or cyclical variation.

The median is the middle value when data is ordered, so it is not affected by extreme values. With right-skewed revenue data, the median remains representative of the typical transaction, making it the robust measure of central tendency. It accurately reflects the center without being distorted by the large enterprise transactions.

Standardised coefficients express each predictor's effect in standard-deviation units, so they are directly comparable across variables measured on different scales. The variable with the largest absolute standardised coefficient, sqft_living, therefore exerts the strongest effect on price.

A p-value of 0.03 falls below the 0.05 significance threshold, so the null hypothesis of equal conversion rates is rejected. The result is statistically significant, indicating the observed difference between the two website designs is unlikely to arise from chance alone.

Mean, median and mode each summarise a dataset with a single representative value, describing where data clusters. They are the standard measures of central tendency in descriptive statistics, distinct from dispersion measures such as range or standard deviation.

The median is the middle value when data is ordered, so it is not influenced by a few extremely large orders. In a right-skewed distribution where mean > median, the median better reflects the typical order amount for the majority of customers, making it the appropriate measure to report.

Recall, also known as sensitivity or true positive rate, measures the proportion of actual positives that are correctly identified. In this scenario, the analyst wants to correctly identify actual churners, so recall directly addresses that goal. High recall means the model captures most churners, which is crucial when the cost of missing a churner is high.

Seasonality describes a repeating pattern that recurs at fixed intervals shorter than a year, such as peak sales every December. This matches the stem's consistent annual December spike, distinguishing it from trend, which is a long-term directional movement.

R² quantifies the share of total variance in the dependent variable accounted for by the fitted regression on the independent variable, expressed as a proportion between 0 and 1. It therefore directly satisfies the stem's requirement to measure explained variance rather than correlation strength or slope.

R² measures the proportion of variance in the dependent variable accounted for by the regression model. A value of 0.85 means the model explains 85% of that variance, leaving 15% attributable to other factors or random error.

To calculate the difference between dates, the string must be converted to a date data type so that date arithmetic can be performed. Many tools and languages provide functions to parse strings into dates, enabling subtraction to yield the number of days.

Recall measures the proportion of actual churners correctly identified, directly matching the stated focus on catching churners. With only 10% churners, accuracy is misleading because predicting 'no churn' everywhere scores 90%, whereas recall exposes missed churners.

Time series analysis explicitly models sequential dependence and seasonality through components such as trend, seasonal, and residual terms, satisfying the stem's clear seasonal pattern requirement. Unlike regression, it uses autocorrelation and prior-period values, so forecasts of future sales account for recurring cycles rather than treating observations as independent.

RFM analysis scores customers on recency, frequency and monetary value derived from purchase history, directly satisfying the segmentation requirement. It groups customers into actionable tiers such as best, loyal or at-risk based on transactional behaviour.

The query groups orders by customer_id, filters order_date to the 2023 range, and applies a HAVING count greater than five. This returns customers whose 2023 order count exceeds five, satisfying the stated purpose of identifying high-frequency customers.

Effect size quantifies the minimum difference the test must detect between variants. Smaller effects demand larger samples, so specifying it directly drives the sample size calculation and satisfies the validity constraint, preventing an underpowered test that misses genuine differences.

ANOVA compares means across three or more independent groups in one test, controlling the family-wise error rate that repeated t-tests would inflate. With three training methods as the grouping factor, it fits the scenario directly.

The Pearson correlation coefficient measures the strength and direction of a linear relationship between two continuous variables. Temperature and ice cream sales are both continuous, and the analyst seeks a linear association, making Pearson correlation the appropriate measure. It ranges from -1 to 1, indicating perfect negative to perfect positive linear relationships.

The desired effect size is the minimum detectable difference between designs that the test must reliably detect. Smaller effects require substantially larger samples, making it a core input to any sample size calculation for the A/B test.

Interpolation estimates missing points from neighbouring known values along the time axis, preserving the daily temperature trend and seasonality. Mean or median imputation flattens local variation, whereas interpolation maintains temporal continuity, satisfying the requirement to keep the series' trend intact.

A mean above the median indicates right-skew, where high-value outliers pull the mean upward while the median stays resistant. This satisfies the stem's constraint of mean ($350,000) exceeding median ($280,000), confirming a positively skewed distribution of house prices.

Min-max normalisation linearly transforms each feature to a fixed 0–1 range using the minimum and maximum values, satisfying the requirement to bound features within a common scale. This suits algorithms sensitive to magnitude, such as k-nearest neighbours or neural networks, where unbounded inputs distort distance calculations.

A Pearson correlation of 0.85 indicates a strong positive linear association. This means that as advertising spend increases, sales revenue tends to increase linearly. It does not imply causation, but it does summarize the strength and direction of the linear relationship. This is the most appropriate conclusion based solely on the correlation coefficient.

Accuracy verifies that values correctly reflect the real-world entities or events they describe, directly satisfying the stem's requirement for data representing reality without errors. Unlike completeness, which checks for missing values, or consistency, which checks formatting agreement across sources, accuracy measures correctness against an authoritative reference.

A histogram bins the numeric 1–5 ratings into adjacent intervals and plots frequency per bin, directly revealing the distribution's shape, spread and skew. Bar charts suit categorical counts, not ordered numeric scales, so they obscure distributional patterns the analyst needs.

Larger samples shrink the standard error of the estimated effect, which directly raises statistical power — the probability of detecting a true difference when one exists. This satisfies the scenario's need to distinguish genuine treatment effects from random variation.

Post-pruning involves growing the full tree and then removing branches that provide little improvement in validation performance. This technique, such as cost-complexity pruning, simplifies the tree and reduces overfitting. It is effective because it uses validation data to guide the pruning process, ensuring that only branches that contribute to predictive accuracy are retained.

A scatter plot displays individual data points on a two-dimensional plane, with one variable on the x-axis and the other on the y-axis. It allows the analyst to visually assess the strength, direction, and form of the relationship between two continuous variables. It is the standard tool for examining correlation and identifying patterns or outliers.

Missing values are often imputed to maintain dataset completeness.

Completeness measures whether all required values are present in a record. A missing phone number means the attribute holds no value, directly violating this dimension. Other dimensions, such as accuracy or consistency, assume a value exists but is wrong or conflicting; here nothing is recorded at all.

One-hot encoding with k-1 categories (here, three) avoids the dummy variable trap by preventing perfect multicollinearity. The dropped category becomes the reference level, and the coefficients for the other categories represent the difference from that reference. This is the standard approach for including nominal categorical variables in linear regression.

K-means clustering partitions unlabelled data into k groups by minimising within-cluster variance, directly satisfying the stem's requirement for segmentation without prior labels. Unlike supervised methods such as logistic regression or decision trees, it needs no target variable, making it the appropriate choice for discovering behavioural customer segments.

Treating 'N/A' as a distinct category preserves the information that the region is missing or not applicable, which can be predictive. This approach avoids introducing false assumptions and allows the model to learn any signal associated with missingness. It is a common and valid strategy for categorical variables.

PCA projects the 1000 features onto orthogonal principal components ordered by explained variance, letting you keep the top components that retain most variance while discarding the rest, reducing dimensionality despite having fewer samples than features.

Applying a distinct count to the user identifier within each date collapses multiple sessions for the same user into one, producing the true number of unique users active that day. This directly matches the definition of daily active users and prevents session volume from distorting the metric.

A dashboard allows multiple visual elements to be combined. A pie chart clearly shows each category's proportion of total revenue, while a separate KPI card prominently displays the overall total. This directly satisfies both requirements: comparing shares and having a distinct panel for the total. It is the most effective and standard approach for this type of reporting need.

The logarithmic transformation is highly effective for right-skewed data with positive values, as it compresses large values more than small ones, reducing skewness and stabilizing variance. It is commonly used for monetary amounts and can make the distribution more symmetric, improving the performance of linear models. It also helps with heteroscedasticity.

Substituting x = 10 into the fitted equation gives y = 2.5 + 1.2(10) = 2.5 + 12 = 14.5. The slope of 1.2 contributes twelve units across the ten-unit increase, and the intercept of 2.5 adds the baseline, yielding the predicted response.

A large gap between training error and validation error is the classic symptom of overfitting, because the model has memorized training noise that does not generalize. Comparing the two error values directly quantifies that gap and tells the analyst whether the model's complexity is justified, making this a core diagnostic step before deployment.

The silhouette score measures how similar each point is to its own cluster versus the nearest other cluster, producing a coefficient between -1 and 1. The cluster count with the highest average silhouette indicates the best-separated segmentation.

ANOVA compares means across three or more groups simultaneously while controlling family-wise error. With normal distributions and roughly equal variances, its assumptions hold, making it correct for comparing average sales across the three store locations.

Logistic regression outputs a calibrated probability of the positive class, so 0.75 means a 75% estimated likelihood of churn for that customer. This satisfies the stem's interpretation requirement, distinguishing probabilistic output from a deterministic classification decision.

Desired effect size is the minimum lift worth detecting between control and variant checkout processes. Smaller effects need larger samples, so effect size is a required input when calculating sample size for the A/B test.

Correcting inconsistent data resolves conflicting formats, units and values so records agree across sources. This is a core data cleaning activity, satisfying the stem's requirement to identify steps that standardise and reconcile raw data before analysis.

The chi-square test of independence compares observed versus expected frequencies across a contingency table, determining whether two categorical variables are associated. Gender and product preference are both nominal categories, so this test satisfies the requirement for assessing significant association between them.

With α=0.05, the slope p-value of 0.03 falls below the threshold, so the slope is statistically significant. R-squared of 0.45 means income explains 45% of the variance in spending. Both conditions in the stem are satisfied simultaneously.

Non-constant variance and non-normal residuals violate linear regression assumptions. A logarithmic transformation of the dependent variable compresses the scale of large values, stabilising variance and pulling the residual distribution towards normality, which is the standard remedy for this pattern of heteroscedasticity.

Validity checks that values conform to a defined domain, format or acceptable range, directly matching the stem's requirement. Unlike accuracy, which concerns correspondence to real-world truth, or consistency, which concerns agreement across sources, validity enforces permitted value sets and boundaries, so out-of-range or malformed entries fail the check.

The elbow marks where additional clusters stop yielding meaningful WCSS reduction. At k=4 the curve flattens, so moving to k=5 or beyond adds complexity without materially lowering within-cluster sum of squares, confirming four clusters as the sensible choice.

The magnitude of r, 0.85, exceeds the conventional 0.7 threshold for strong association, and the negative sign indicates that as advertising spend rises, sales tend to fall along a linear trend. Direction and strength are therefore both captured by this interpretation.

Recall, also known as sensitivity, is the proportion of actual positives that are correctly identified, calculated as TP/(TP+FN) = 100/130 ≈ 0.769. It measures the model's ability to find all actual churners, which aligns with the analyst's goal. A high recall means few churners are missed, which is often critical in churn prediction to avoid losing customers.

Overfitting occurs when a model learns noise and idiosyncrasies in the training set rather than generalisable patterns, producing the 99% versus 65% gap. The stem's constraint — high training accuracy with poor unseen-data performance — is the defining signature of variance-dominated overfitting, so regularisation, pruning or more data would be required.

Imputing with the mean replaces each missing 'age' entry with the column's arithmetic average, preserving the existing sample size and keeping the dataset's overall mean unchanged. This satisfies the stem's requirement for a valid numerical handling method, since the mean is calculable only on quantitative data such as age.

Pruning reduces the size of the tree by removing branches that provide little predictive power, which helps combat overfitting. By simplifying the model, pruning improves generalization to unseen data. In this scenario, the large gap between training and test accuracy indicates overfitting, so pruning is the appropriate technique to enhance test performance.

The median is positional, not arithmetic: it depends only on the middle value's rank, so extreme outliers cannot drag it. The mean incorporates every value's magnitude, making it highly sensitive to skew. For skewed data, the median better represents typical central tendency.

For a normally distributed variable, the mean preserves the central tendency and keeps the overall distribution shape intact, unlike median or mode imputation. Since the data is symmetric, the mean is the best unbiased estimate for replacing missing numeric entries, satisfying the normality constraint in the stem.

The median depends only on positional order, so extreme values shift it minimally. The mean sums every value and is dragged strongly by outliers, while the mode reflects only the most frequent value and can be unstable in continuous data.

The elbow method selects the k where the within-cluster sum of squares drops sharply before levelling off. The sharp bend at k=3 marks that inflection point, so three clusters is recommended; the minor bend at k=5 is a secondary, weaker signal.

Pruning involves growing a full tree and then removing branches that provide little predictive power, reducing complexity and overfitting. This technique, such as cost-complexity pruning, balances bias and variance, improving generalization. For churn classification, pruning helps avoid capturing noise in the training data, leading to more robust predictions on new customers.

Normality is assumed for the t-test, though it is robust for large samples.