Temporal and Multivariate Visualization

Introduction

Many analytical questions are inherently temporal or multivariate. A production manager wants to know not only whether defects increased, but when they increased, which product lines were affected, and whether the change followed a supplier or process intervention. A financial analyst must compare prices, volatility, trading volume, and macroeconomic indicators across different time scales. A public-health researcher may examine cases, mobility, vaccination, demographic characteristics, and geographic variation simultaneously.

Temporal and multivariate visualization provides methods for representing these relationships while preserving analytical integrity. Temporal visualization concerns observations ordered in time, including regularly sampled time series, irregular events, intervals, durations, and repeated cycles. Multivariate visualization concerns observations described by several variables, often with mixed measurement scales and complex dependencies. The two areas frequently overlap: a time series may contain dozens of variables, and a multivariate data set may be investigated through time.

The central difficulty is not displaying more marks. It is enabling valid perception and comparison without concealing uncertainty, scale differences, missingness, aggregation effects, or important interactions. This chapter develops a design framework based on transformations, small multiples, dimensionality reduction, and coordinated views. It also addresses temporal aggregation, event representation, high-dimensional structure, interaction, evaluation, ethics, and implementation.

Learning Objectives

After completing this chapter, a student should be able to:

1. Distinguish temporal data types, temporal granularities, and event structures.

2. Select transformations and visual encodings appropriate to temporal and multivariate data.

3. Design and critique small-multiple displays.

4. Explain the assumptions and interpretation of principal component analysis and related dimensionality-reduction methods.

5. Construct coordinated multiple views using linking and brushing.

6. Diagnose misleading effects caused by aggregation, normalization, missing data, and overplotting.

7. Apply these methods to a realistic professional problem while considering uncertainty, privacy, and governance.

1. Temporal Data: Structure and Semantics

A temporal data set consists of observations associated with time. Time may be represented as an instant, interval, duration, sequence, or recurring calendar position. These distinctions matter because an event occurring at 09:00 is not equivalent to an activity lasting from 09:00 to 11:00, and a monthly total is not necessarily comparable with a daily measurement.

1.1 Time series, events, and intervals

A time series is an ordered sequence of measurements, commonly written as \(y_t\), where \(t\) denotes an ordered time index. The observations may be regularly spaced, such as hourly electricity readings, or irregularly spaced, such as machine-failure reports. A multivariate time series can be written as \(\mathbf{x}_t=(x_{1t},x_{2t},\ldots,x_{pt})\), where \(p\) variables are recorded at time \(t\).

Event data records occurrences such as transactions, alarms, or hospital admissions. Events may have attributes including category, location, severity, and actor. Interval data records a start and end time, allowing duration and overlap to be analyzed. A timeline, Gantt chart, interval plot, or calendar view may be more appropriate for intervals than a conventional line chart.

Time has semantic properties beyond ordering. Calendar time contains months, weekdays, holidays, and time zones. Duration time measures elapsed time. Business time excludes weekends or uses organization-specific calendars. A visualization should state which interpretation is used. Misalignment of time zones, daylight-saving changes, or inconsistent timestamp precision can generate apparent patterns that are data-processing artifacts.

1.2 Components of temporal behavior

A useful descriptive decomposition of a series is:

\[

y_t = T_t + S_t + R_t

\]

for an additive representation, where \(T_t\) is trend, \(S_t\) is seasonal or recurring structure, and \(R_t\) is the remainder. This is a descriptive model, not a universal law. It is appropriate when the magnitudes of seasonal fluctuations are approximately constant across the level of the series. When seasonal variation grows with the level, a multiplicative form is often more suitable:

\[

y_t = T_t S_t R_t.

\]

Multiplicative models require compatible nonnegative quantities and are commonly analyzed on a logarithmic scale, because \(\log(y_t)=\log(T_t)+\log(S_t)+\log(R_t)\) when all terms are positive.

Visualization should not imply that a decomposition has been statistically established merely because a smooth curve has been drawn. Smoothing can reveal broad structure but can also erase shocks, shift apparent turning points, and create boundary artifacts. A caption should identify the smoothing method, window or bandwidth, and whether values are centered or trailing.

1.3 Temporal transformations

Transformations can make a pattern visible, but each changes the question being asked. Common operations include:

• Resampling: converting observations to a new frequency, such as hourly to daily.
• Aggregation: computing sums, means, medians, minima, maxima, or counts within time bins.
• Rolling summaries: calculating a statistic over a moving window.
• Differencing: examining change, such as \(\Delta y_t=y_t-y_{t-1}\).
• Growth rates: comparing relative change, commonly \((y_t-y_{t-1})/y_{t-1}\) when the denominator is nonzero.
• Indexing: setting a reference value to 100, so \(I_t=100y_t/y_0\), where \(y_0\) is the chosen baseline.
• Lagging: comparing a variable with its earlier value, such as \(y_{t-k}\).
• Log transformation: using \(\log(y_t)\) for positive, right-skewed values.

A sum is appropriate for additive quantities such as total energy consumed, whereas a mean may be appropriate for temperature. Averaging rates can be misleading when denominators differ; a weighted rate or reconstruction from numerators and denominators may be required. For example, an overall defect rate should generally be total defective units divided by total inspected units, not the unweighted mean of plant-level rates.

2. Designing Temporal Displays

2.1 Encoding time

Time is usually encoded by horizontal position because position supports accurate comparison. Lines connect ordered measurements and are useful for continuous or regularly sampled processes. Points should be retained when observations are sparse, irregular, or uncertain. Area charts can emphasize cumulative quantity, but stacked areas make individual components difficult to compare because only the baseline series has a stable reference. Bars are effective for discrete periods, especially when zero-based magnitude comparison is important.

A temporal display should make the following explicit:

• the time zone and calendar convention;
• the temporal resolution and aggregation rule;
• whether intervals represent point observations or durations;
• the treatment of missing observations;
• the meaning of line connections;
• the baseline and units.

Dual axes are particularly risky. Two scales can make unrelated variables appear correlated or can conceal a scale choice that changes the visual conclusion. A shared normalized scale, separate aligned panels, or an indexed representation is usually more transparent.

2.2 Seasonality, cycles, and calendar effects

Seasonal behavior should be investigated at the relevant period. A line chart across several years may show annual cycles, but a seasonal subseries plot can compare January values across years, February values across years, and so forth. Heat maps with time on one axis and calendar position on the other are effective for identifying recurring patterns, such as weekday-hour demand.

The choice of aggregation can create or remove seasonality. Monthly totals may hide intraday peaks; daily averages may hide weekend effects. Analysts should examine multiple granularities and preserve the original data for auditability. A calendar heat map is useful for scanning, but color intensity should not be the sole basis for precise comparisons; tooltips, labels, or linked detail views can provide exact values.

2.3 Events and narratives

Events can be encoded as points, vertical rules, bands, or annotations. A vertical rule is suitable for a single intervention date. A shaded band is suitable for a policy period or maintenance interval. Event labels should be selective: excessive annotation produces clutter and encourages post hoc storytelling.

For intervention analysis, a pre/post display can be informative, but visual coincidence does not prove causation. Other changes may occur at the same time, and a trend may have changed without the intervention. A credible analysis may require a comparison group, interrupted time-series model, or other causal design. Visualization should therefore display control series, uncertainty, and relevant alternative explanations where possible.

3. Multivariate Data and Visual Encoding

Multivariate observations have several measured attributes. Variables may be quantitative, ordinal, nominal, binary, compositional, or textual. Their types determine valid operations and encodings. A mean is meaningful for a quantitative variable but not generally for a nominal category. Color hue is effective for a small number of unordered categories; ordered magnitude is better represented by position, length, or a perceptually ordered color scale.

3.1 Common multivariate displays

A scatterplot shows the relationship between two quantitative variables. Adding color, shape, size, or facets can introduce additional variables, but each channel has limited capacity. Size is less precise than position, and simultaneous use of size and color can overload interpretation. A scatterplot matrix supports pairwise exploration but becomes large as the number of variables grows.

A parallel-coordinates plot represents each variable as a vertical axis and each observation as a polyline crossing those axes. It can reveal clusters and profiles, but scale normalization, axis ordering, and overplotting strongly affect interpretation. Axes should be reordered deliberately, and brushing or transparency is often necessary.

A correlation heat map summarizes pairwise association. For Pearson correlation, the coefficient between variables \(X\) and \(Y\) is

\[

r=\frac{\sum_{i=1}^{n}(x_i-\bar{x})(y_i-\bar{y})}{\sqrt{\sum_{i=1}^{n}(x_i-\bar{x})^2}\sqrt{\sum_{i=1}^{n}(y_i-\bar{y})^2}}.

\]

Here, \(n\) is the number of paired observations and \(\bar{x},\bar{y}\) are sample means. Pearson correlation measures linear association and is sensitive to outliers; it does not establish causation. A heat map should communicate sample size, missing-data treatment, and whether correlations were computed across observations, time points, or groups.

A table remains valuable when exact values, ranks, or auditability matter. Visualization should complement, not automatically replace, tabular evidence.

3.2 Normalization and comparability

Variables measured in different units cannot be compared directly in many displays. Standardization uses

\[

z_i=\frac{x_i-\bar{x}}{s},

\]

where \(s\) is the sample standard deviation, assuming a nonzero standard deviation. The resulting value expresses how far an observation is from its variable’s mean in standard-deviation units. Standardization is sensitive to outliers and can obscure meaningful absolute differences. Min–max scaling, robust scaling based on median and interquartile range, or domain-specific normalization may be preferable.

Normalization is not neutral. Per-capita rates, percentages, indexed values, and z-scores answer different questions. Every transformed display should identify its reference population, denominator, baseline, and treatment of zero or negative values.

4. Small Multiples and Structured Comparison

Small multiples repeat a consistent graphical design across categories, locations, variables, or time periods. Their power comes from aligned scales, repeated visual grammar, and reduced occlusion. For example, identical line charts for twenty factories allow a reader to compare trends without forcing twenty lines into one panel.

A valid small-multiple design requires careful decisions:

1. Facet variable: choose the dimension that supports the intended comparison.

2. Common scales: use shared axes when magnitude comparison matters.

3. Free scales: use independent scales only when revealing within-panel shape is more important, and clearly signal that magnitudes are not directly comparable.

4. Ordering: sort panels by geography, magnitude, trend, or an analytically justified variable—not by an accidental display order.

5. Reference structure: retain common baselines, event markers, and units.

6. Resolution: ensure labels and marks remain legible at the intended viewing size.

A common-scale display can make small entities appear visually flat, while free scales can exaggerate differences in shape. One solution is a two-stage design: a common-scale overview followed by selected free-scale detail. Another is to include a reference line, range indicator, or inset showing absolute magnitude.

Small multiples also support comparison of multivariate profiles. A matrix of scatterplots, box plots for several variables, or repeated maps can reveal heterogeneous relationships. Their limitation is the cognitive cost of scanning many panels and remembering positions. Direct annotations, consistent ordering, and linked interaction reduce this cost.

5. Dimensionality Reduction

When variables are numerous, direct displays become crowded. Dimensionality reduction maps high-dimensional observations into a lower-dimensional representation, commonly two or three dimensions. It is primarily an exploratory technique: proximity in the display is meaningful only under the method, scaling, distance metric, and preprocessing choices used.

5.1 Principal component analysis

Principal component analysis (PCA) finds orthogonal directions that successively maximize variance. Let \(X\) be an \(n\times p\) data matrix after appropriate centering, and possibly standardization. The first principal direction \(\mathbf{w}_1\) solves

\[

\max_{\|\mathbf{w}\|=1} \operatorname{Var}(X\mathbf{w}).

\]

The constraint \(\|\mathbf{w}\|=1\) prevents arbitrary rescaling. Subsequent directions are constrained to be orthogonal to earlier directions. In practice, PCA can be computed through the eigendecomposition of the covariance or correlation matrix, or through singular value decomposition.

A principal component score is the projection \(\mathbf{z}_k=X\mathbf{w}_k\). The loading vector \(\mathbf{w}_k\) indicates how original variables contribute to component \(k\), subject to sign indeterminacy: reversing all signs produces the same geometry. The proportion of variance explained by component \(k\) is

\[

\frac{\lambda_k}{\sum_{j=1}^{p}\lambda_j},

\]

where \(\lambda_k\) is the corresponding eigenvalue. This proportion describes variance captured, not predictive accuracy, causal importance, or semantic meaning.

PCA assumes quantitative variables and is affected by units, scale, outliers, and missing-data treatment. Standardization is often appropriate when variables have incomparable units, but not when units themselves convey a meaningful weighting. PCA captures linear structure and may fail on curved manifolds or categorical data. A biplot can show scores and loadings, but it should not be interpreted as a precise map of all original distances.

5.2 Other methods

Multidimensional scaling (MDS) seeks a low-dimensional configuration preserving selected pairwise dissimilarities. It is useful when a meaningful distance matrix is available. t-distributed stochastic neighbor embedding (t-SNE) emphasizes local neighborhoods and is useful for exploratory cluster discovery, but its global distances, cluster sizes, and results across runs should not be overinterpreted. Uniform Manifold Approximation and Projection (UMAP) also emphasizes neighborhood structure and depends on metric and parameter choices.

For categorical or mixed data, methods such as multiple correspondence analysis, factor analysis of mixed data, or distance-based approaches may be more suitable than ordinary PCA. For temporal data, dynamic embedding, trajectory visualization, and state-space representations can preserve sequential structure that a static projection loses.

A responsible workflow compares several parameter settings, checks stability, overlays known labels only after unsupervised exploration, and links projected points back to original variables. Dimensionality reduction is a view of the data, not a replacement for the data.

6. Coordinated Multiple Views

A coordinated multiple-view system assigns related analytical tasks to separate but linked displays. A dashboard might include a time-series overview, a geographic map, a scatterplot, and a detail table. Selecting or brushing records in one view highlights the same records elsewhere. Filtering changes the active subset, while zooming changes scale without necessarily changing membership.

Coordination supports the information-seeking cycle of overview, zoom, filter, details-on-demand, and relate. Effective systems make state visible: selected records, active filters, date ranges, and transformations should be clearly indicated. Reset controls and reversible actions are essential.

Linking is most useful when views encode complementary questions. A PCA scatterplot can identify a group of unusual machines; a linked time chart can show when their behavior changed; a parallel-coordinates view can reveal which sensors distinguish them. However, linked views may amplify selection bias. Analysts can repeatedly select interesting subsets until an appealing pattern appears. Systems should support saved selections, audit logs, reproducible queries, and explicit distinction between exploratory findings and confirmatory results.

7. Worked Example: Monitoring a Distributed Data-Centre Fleet

Suppose an IT operations team monitors 120 servers at five-minute intervals for six months. Variables include CPU utilization, memory utilization, disk latency, network throughput, fan speed, temperature, and incident status. The objective is to identify early indicators of thermal incidents without confusing workload cycles with failures.

Step 1: Temporal quality assessment

The analyst first verifies timestamp time zones, duplicate records, sensor calibration, missing intervals, and server commissioning dates. A missing reading is not automatically zero. Missingness is displayed as gaps or a separate status layer, while imputed values, if used, are marked explicitly.

Step 2: Multi-scale temporal views

A line chart of fleet-level median temperature gives an overview, while hourly and daily summaries reveal operational cycles. Because temperature is a measurement, a mean or median may be appropriate; a sum would have no meaningful interpretation. A heat map with server on the vertical axis and time on the horizontal axis reveals whether incidents are isolated or fleet-wide.

Step 3: Derived variables

For each server, the team computes a rolling 30-minute median and a change from the previous hour. These transformations are labeled as derived features. A normalized indexed chart compares thermal trajectories relative to each server’s first valid week, while the raw-temperature chart remains available because an index can conceal absolute risk.

Step 4: Small multiples

Small multiples show temperature and CPU utilization for each server, with a common y-axis and incident markers. Servers are ordered by incident count, but the order is accompanied by an explanation. A second view focuses on the ten servers with the largest sustained temperature increase. This avoids displaying 120 overlapping lines.

Step 5: Multivariate analysis

The team standardizes sensor variables across servers and applies PCA for exploratory analysis. The first two components explain 61% of standardized variance in this illustrative data set. A score plot suggests a group characterized by high temperature, fan speed, and disk latency. Loadings are inspected before assigning meaning to the axes. The team does not label the group “at risk” solely because it forms a visual cluster.

Step 6: Coordinated investigation

Brushing the suspected group in the PCA plot highlights corresponding servers in the heat map and time-series view. The analysts observe that the pattern occurs mainly during a recent firmware rollout and is concentrated in one hardware model. A linked table shows that the affected servers share a rack-level cooling configuration.

Outcome

The visualization does not by itself prove that firmware caused the incidents. It narrows the investigation and motivates controlled testing. Engineers compare affected and unaffected servers, inspect firmware logs, and measure rack airflow. The eventual operational response may include rollback, cooling adjustment, and enhanced monitoring.

Learning points

The case demonstrates that temporal aggregation, multivariate scaling, dimensionality reduction, and coordination answer different questions. Raw values preserve operational meaning; normalized values support comparison; PCA summarizes covariance; linked views support traceability. The workflow succeeds because every transformation remains inspectable and because visual evidence is treated as a basis for investigation rather than proof of causation.

8. Evaluation, Limitations, and Governance

8.1 Analytical limitations

Temporal displays can suffer from aliasing: a sampling frequency that is too low may miss short events. Aggregation can produce ecological fallacies, in which group-level patterns are incorrectly attributed to individuals or machines. Simpson’s paradox can cause an aggregate trend to differ from subgroup trends. Analysts should therefore inspect relevant subgroups and retain denominators.

Overplotting, excessive color categories, distorted axes, and unstable layouts reduce readability. Missing data may be informative rather than random; sensor failure during high load may itself be a warning signal. Outliers require investigation, not automatic deletion. Confidence bands and prediction intervals should be distinguished from standard-error bands, and uncertainty should not be represented as if it were measurement certainty.

Dimensionality-reduction plots are especially vulnerable to overinterpretation. A visually separated cluster may result from scaling, parameter selection, or a categorical variable encoded in the preprocessing. Repeated analyses with alternative preprocessing and validation on independent data are essential for high-stakes conclusions.

8.2 Ethical and regulatory considerations

Temporal and multivariate data can reveal individual behavior, health status, location, or employment performance. Fine-grained timestamps may permit re-identification even when names are removed. Governance should address purpose limitation, access control, retention, consent or lawful basis, aggregation thresholds, and disclosure risk. Dashboards should avoid exposing unnecessary individual-level detail.

Fairness must also be examined. A model or visualization based on unevenly measured groups can make some populations appear more variable or more risky simply because they are monitored more intensely. Missingness, sampling design, and proxy variables should be documented. Accessibility requires adequate contrast, non-color encodings, keyboard support where applicable, readable labels, and alternatives for color-vision deficiencies.

8.3 AI-assisted visualization

AI tools can recommend chart types, detect anomalies, summarize clusters, or generate natural-language explanations. These capabilities are useful for exploration but can produce unsupported narratives, inappropriate transformations, or hidden selection criteria. Professional practice requires provenance: preserve the data version, transformation code, model parameters, generated recommendations, and human decisions. AI-generated interpretations should be treated as hypotheses subject to domain review, not as evidence.

9. Contemporary Practice and Best-Practice Framework

Modern visualization systems increasingly combine streaming data, uncertainty displays, semantic zooming, notebook-based reproducibility, and web-based interaction. Real-time systems must distinguish latency from absence of data and indicate whether the latest point is complete. Progressive rendering can improve responsiveness, but preliminary values should be labeled.

A robust workflow is:

1. Define the decision and the unit of analysis.

2. Identify temporal semantics, variable types, denominators, and data provenance.

3. Inspect raw observations before transforming them.

4. Choose encodings based on perceptual accuracy and comparison goals.

5. Use aggregation and normalization only with explicit justification.

6. Provide overview and detail, preferably through small multiples or coordinated views.

7. Use dimensionality reduction for exploration, with sensitivity checks and linked original variables.

8. Display uncertainty, missingness, event definitions, and sample sizes.

9. Test the visualization with representative users and accessibility checks.

10. Preserve reproducible specifications, transformation logs, and interpretation boundaries.

A successful advanced visualization is not the most elaborate display. It is a transparent analytical instrument that allows a viewer to compare, investigate, question, and trace conclusions back to data and assumptions.

Chapter Summary

Temporal visualization represents ordered observations, events, intervals, cycles, and durations. The choice of line, point, bar, heat-map, interval, or annotation display should follow the temporal semantics and analytical task. Resampling, aggregation, rolling summaries, differencing, indexing, and logarithmic transformation can reveal structure, but each changes interpretation and must be documented.

Multivariate visualization requires attention to variable type, scale, perceptual encoding, and association. Small multiples provide structured comparison by repeating a consistent design, while coordinated multiple views connect complementary analytical perspectives through selection, filtering, and linking. Dimensionality-reduction methods such as PCA, MDS, t-SNE, and UMAP help explore high-dimensional structure, but their results depend on preprocessing, metrics, parameters, and assumptions. PCA summarizes variance rather than causation or predictive importance.

Professional practice requires checking missingness, aggregation effects, outliers, uncertainty, sampling bias, and privacy risk. In the data-centre case, linked temporal and multivariate views narrowed an operational investigation without incorrectly claiming causality. The most reliable systems preserve raw context, expose transformations, support accessibility, and maintain reproducible analytical provenance.

Key Terms

Temporal data: Data associated with ordered time points, intervals, or durations.

Time series: Sequential observations indexed by time.

Event data: Records of occurrences with times and attributes.

Interval data: Records with a start time and an end time.

Seasonality: Repeating variation associated with a calendar or recurring period.

Resampling: Converting observations to a different temporal frequency.

Rolling summary: A statistic computed over a moving time window.

Differencing: Subtracting an earlier observation from a later observation.

Indexing: Expressing values relative to a selected baseline.

Small multiples: Repeated, consistently formatted charts across facets.

Multivariate data: Observations containing multiple variables.

Normalization: Transforming values to support comparison across scales or units.

Correlation: A measure of association between variables; Pearson correlation measures linear association.

Parallel coordinates: A display representing observations as lines crossing parallel variable axes.

Dimensionality reduction: Mapping high-dimensional observations into fewer dimensions.

Principal component analysis: A method finding orthogonal directions that successively maximize variance.

Loading: The coefficient indicating a variable’s contribution to a principal component.

Brushing and linking: Selecting records in one view and highlighting them in related views.

Overplotting: Occlusion caused by too many marks occupying the same display region.

Provenance: A record of data sources, transformations, parameters, and analytical decisions.

Review Questions

Short Answer Questions

1. Distinguish a time series, an event, and an interval data record.

2. Why can aggregation change the interpretation of a temporal pattern?

3. What is the principal purpose of small multiples?

4. What does a PCA loading represent?

5. Why should missing observations not automatically be encoded as zero?

Long Answer Questions

1. Explain how trend, seasonality, and remainder can be represented in temporal analysis. Discuss when additive and multiplicative representations are appropriate.

2. Compare line charts, heat maps, calendar views, and interval charts for temporal data. Include the conditions under which each is most useful.

3. Explain PCA mathematically and critically evaluate the effects of centering, standardization, outliers, and variable units.

4. Discuss how coordinated multiple views support exploratory analysis and identify risks of interactive selection.

5. Develop a design strategy for a multivariate operational dashboard that includes uncertainty, accessibility, governance, and reproducibility.

Analytical / Application Questions

1. A company reports average defect rates for four plants but gives each plant equal weight despite very different production volumes. Explain why the aggregate rate may be misleading and propose a better visualization and calculation strategy.

2. Design a coordinated visualization for investigating whether a software release changed server performance. Specify the temporal transformations, small multiples, multivariate view, event annotations, and evidence needed before claiming causation.

3. A PCA projection shows three visually distinct groups, but the groups disappear after robust scaling and removal of one outlier. Analyze what this implies about the original conclusion and propose validation steps.

Multiple Choice Questions with Answers

Question 1

Which transformation expresses each value relative to a selected baseline value of 100?

A. Differencing

B. Indexing

C. Winsorization

D. Binning

Correct Answer: B. Indexing

Question 2

When is an additive temporal decomposition generally most appropriate?

A. When seasonal fluctuations grow proportionally with the series level

B. When all observations are categorical

C. When seasonal fluctuation magnitudes are approximately constant

D. When the series contains no time ordering

Correct Answer: C. When seasonal fluctuation magnitudes are approximately constant

Question 3

What is the principal purpose of small multiples?

A. To compare repeated displays using a consistent visual design

B. To hide variation between categories

C. To replace all numerical tables

D. To guarantee causal conclusions

Correct Answer: A. To compare repeated displays using a consistent visual design

Question 4

In PCA, the first principal direction is selected to maximize what quantity subject to a unit-length constraint?

A. The number of missing values

B. Projected variance

C. The number of categories

D. The median absolute error

Correct Answer: B. Projected variance

Question 5

What does a Pearson correlation coefficient primarily measure?

A. Linear association between paired quantitative variables

B. Causal impact of one variable on another

C. The number of clusters in a data set

D. The duration of an event

Correct Answer: A. Linear association between paired quantitative variables

Question 6

Why is standardization often used before comparing variables in PCA?

A. It converts all variables into categorical data

B. It gives variables measured on different scales comparable influence

C. It removes all outliers automatically

D. It guarantees normally distributed variables

Correct Answer: B. It gives variables measured on different scales comparable influence

Question 7

Which interaction highlights the same selected records across multiple coordinated views?

A. Brushing and linking

B. Aliasing

C. Rasterization

D. Downsampling only

Correct Answer: A. Brushing and linking

Question 8

Which statement about t-SNE or UMAP is most appropriate?

A. They prove that displayed clusters are causal groups

B. Their global distances and cluster sizes should always be interpreted literally

C. They are exploratory methods whose results depend on settings and preprocessing

D. They require every variable to have identical units

Correct Answer: C. They are exploratory methods whose results depend on settings and preprocessing

Question 9

For total defects across plants with different inspection volumes, which rate is generally most appropriate?

A. The unweighted mean of plant defect rates

B. Total defective units divided by total inspected units

C. The highest plant defect rate

D. The median plant rate regardless of volume

Correct Answer: B. Total defective units divided by total inspected units

Question 10

Which practice best supports responsible temporal and multivariate visualization?

A. Treating missing readings as zero without documentation

B. Using dual axes whenever variables have different units

C. Recording transformations, parameters, and data provenance

D. Interpreting a visual cluster as proof of causation

Correct Answer: C. Recording transformations, parameters, and data provenance