Choosing the Right Chart for Multimodal Data Analysis and Avoiding Misleading Visuals

Chart selection is matching a visualization's type to the analytical question being asked, so the picture's geometry — not its caption — carries the answer. [GROUND TRUTH] (Sources/nca-genm/domain-2-data-analysis.md) provides the canonical mapping directly, as a table: "Distribution of one variable → Histogram, box plot. Relationship between two variables → Scatter plot. Compare categories → Bar chart (column for vertical). Change over time → Line chart. Correlation across many variables → Heatmap. Part-to-whole → Stacked bar (pie sparingly)." Six questions, six answers, and the exam's version of this topic is largely a test of whether you have that six-way mapping memorized cold enough to apply it to a described scenario rather than a labeled chart.

The underlying reason the mapping works is that every chart type encodes data using a specific visual channel — position along an axis, length of a bar, color intensity in a cell, angle of a slice — and each channel is better suited to some comparisons than others. A histogram's bars encode count within a bin, which is exactly what "distribution" means and nothing else. A bar chart's bar length encodes magnitude, which is exactly what "compare categories" needs and a histogram, with its fixed binning logic, does not supply cleanly. Learning the mapping by rote gets you through most exam questions; understanding why each chart's channel fits its question is what lets you handle a scenario worded unusually.

L1 — Intuition

Before matching a scenario to a chart, it helps to hold six short identities in mind, each phrased as the question the chart answers rather than as a description of what it looks like — because the exam's scenario wording asks the question, not the appearance.

L2 — Mechanism

A histogram answers "what does the distribution of this one variable look like" by binning a continuous variable into equal-width intervals and showing the count in each bin as a bar's height, revealing shape, skew, and modality in one glance. A box plot answers a related but narrower question — "how spread out is this variable, and are there outliers" — using five landmarks (minimum, 25th percentile, median, 75th percentile, maximum) plus flagged outliers beyond the IQR fences from M2-01, and its real advantage over a histogram is that several box plots line up side by side cleanly, which makes it the tool of choice for comparing a distribution across categories (say, image file size by source dataset) rather than looking at one distribution in isolation.

A scatter plot answers "are these two variables related, and how" by placing one variable on each axis and letting every row become a single point; the shape the points trace out — a tight line, a loose cloud, a curve, distinct clusters — is exactly the visual evidence M2-02 used to catch a non-linear relationship that Pearson r alone would miss. A bar chart answers "how do these categories compare on this value," with each category getting one bar whose length or height encodes its value — the chart's entire job is magnitude comparison across discrete groups, and it does that job better than almost anything else because bar length is one of the most accurately perceived visual channels available.

A line chart answers "how is this changing over time," connecting ordered observations with a continuous line so a trend, a seasonal pattern, or a sudden change point is visible as the line's shape rather than as a table of numbers you would have to read sequentially to notice the same thing. A heatmap answers "what is the structure of this whole matrix at once," most commonly a correlation matrix across many variables, using color intensity in each cell to encode a value — the chart's real advantage is that it scales to dozens of variable pairs at once, where a page of individual scatter plots would be unreadable.

L3 — The exam-relevant edge case

[GROUND TRUTH] (Sources/nca-genm/domain-2-data-analysis.md) adds a seventh, less-emphasized entry: "Part-to-whole → Stacked bar (pie sparingly)." The qualifier "sparingly" is doing real work and is itself testable: pie charts are the one chart type the source material flags as generally disfavored, because human perception of angle and area is measurably worse than perception of length or position — the same reason a bar chart beats a pie chart for the identical part-to-whole data whenever more than two or three slices are involved. A stacked bar chart encodes the same part-to-whole relationship using length instead of angle, and is the source material's preferred alternative for exactly that reason.

L1 — Intuition

A chart can pick the objectively correct type for its question and still misrepresent the data, because chart type and honest scaling are two separate decisions. This section covers the second one.

L2 — Mechanism

[GROUND TRUTH] (Sources/nca-genm/domain-2-data-analysis.md) names the failure modes directly: "truncated/zoomed axes that exaggerate differences, the wrong chart type, overloaded charts, or 3-D effects that distort values." A truncated or zoomed axis starts a numeric axis somewhere other than a natural baseline (usually zero for a bar chart), which stretches the visual distance between values that are, in the real numbers, close together — a bar chart of two products' sales, 98 and 102 units, drawn with a y-axis running from 90 to 105 instead of 0 to 105, makes what is roughly a 4% difference look like one bar is nearly double the other. The underlying numbers are not falsified anywhere in this manipulation; only the reader's visual impression of their relative size is.

The wrong chart type is choosing a chart whose visual channel does not match the analytical question — using a pie chart for a trend over time, or a line chart connecting unordered categories, which implies an ordering and a continuity between category values that does not exist. Overloaded charts cram too many variables, categories, or series into a single visual, past the point where any one comparison remains legible — a bar chart with forty bars in forty different colors technically contains the data but has abandoned the "readable from the shape" property that makes a chart useful over a table in the first place. 3-D effects add a depth dimension to a chart type that does not need one — a 3-D pie chart or a 3-D bar chart — and the added perspective systematically distorts the apparent size of elements based on their position (nearer slices or bars look larger regardless of their real value), on top of whatever perceptual disadvantage the base chart type already carried.

L3 — The exam-relevant edge case

The source material's closing framing generalizes the whole section into one design principle: "Simple charts suit general audiences; more complex ones suit expert audiences." [GROUND TRUTH] (Sources/nca-genm/domain-2-data-analysis.md). This is a genuinely separate judgment from the honesty question — a heatmap of a fifty-variable correlation matrix is not misleading, but it is the wrong choice for a general audience regardless of how honestly it is drawn, because reading it correctly requires domain fluency a general audience does not have. Complexity-for-the-audience and honesty-of-scaling are two independent axes a chart can fail on, and an exam scenario naming a confused or misled audience could be pointing at either one, depending on the specific complaint described.

Treat the following as a constructed set of scenarios built to make the mapping exercise concrete, not measurements from a real analysis.

Scenario A: "What is the typical file size of the 40,000 images in our
  training corpus, and how much does it vary?"
  -> One continuous variable, distribution question -> HISTOGRAM
     (or box plot if the real ask is spread/outliers specifically)

Scenario B: "Does caption length predict how long the model takes to
  process a training batch?"
  -> Two continuous variables, relationship question -> SCATTER PLOT

Scenario C: "How many training examples do we have for each of our
  seven object-detection classes?"
  -> Categorical vs. count, comparison question -> BAR CHART

Scenario D: "How has our model's validation accuracy changed across
  the last 30 training checkpoints?"
  -> Ordered/time variable vs. numeric -> LINE CHART

Scenario E: "Across our twelve numeric image-quality features (sharpness,
  brightness, contrast, noise level...), which pairs move together?"
  -> Many variables, correlation-structure question -> HEATMAP

Scenario F: "Of our total training budget, what share went to compute,
  storage, and annotation labor?"
  -> Part-to-whole, three categories -> STACKED BAR
     (a pie chart with only 3 slices is defensible but stacked bar is
     the source material's stated preference)

Each scenario's wording contains the cue: "typical" and "vary" point at distribution, "predict" and "does X relate to Y" point at relationship, "how many... for each" points at category comparison, "changed across" points at time, "which pairs move together" points at many-variable correlation, and "share of a total" points at part-to-whole. Learning to hear these cues in a differently-worded exam stem is the transferable skill underneath the six-way table.

Treat the following as three constructed chart descriptions, illustrative rather than measurements from any real dashboard.

Chart 1: A bar chart comparing "Model A accuracy: 89.2%" against
  "Model B accuracy: 91.0%", with the y-axis running from 88% to 92%.
  Diagnosis: truncated axis. The real gap is 1.8 percentage points, but
  because the visible axis range is only 4 points wide, Model B's bar
  looks roughly 45% taller than Model A's -- a wildly exaggerated visual
  impression of a genuinely modest difference.
  Fix: run the y-axis from 0% to 100%, or explicitly annotate the
  truncation and the real numeric gap if space is genuinely limited.

Chart 2: A 3-D pie chart showing training-time share across four GPU
  types, viewed at an angle so the front slices appear larger.
  Diagnosis: 3-D effect distorting an already perceptually weak chart
  type. Angle-based encoding (pie) is already harder to compare
  accurately than length-based encoding (bar); adding a 3-D perspective
  compounds the distortion by making foreground slices look larger
  purely due to camera angle, independent of their real share.
  Fix: a plain 2-D stacked bar chart of the same four categories.

Chart 3: A single dashboard panel overlaying 15 differently-colored line
  series (one per experiment run) on one chart, tracking loss over
  10,000 training steps.
  Diagnosis: overloaded chart. Technically every series is present and
  correctly plotted, but 15 overlapping lines in similar colors make it
  impossible to trace any single run's trajectory or reliably compare
  two specific runs against each other.
  Fix: either facet into several smaller charts (5 runs each), or reduce
  to the 3-4 runs that are actually being compared in this analysis,
  with the rest available on demand rather than all drawn at once.

Each diagnosis names the specific mechanism — an exaggerated visual gap, a compounded perceptual distortion, an illegible density of series — rather than a vague "this chart looks bad," because the exam's version of this question asks what specifically is wrong, and "wrong chart type" versus "misleading scaling" versus "overloaded" are three different keyed answers to three differently-worded stems.

⭐ THE EARNED INSIGHT

A chart can be drawn from perfectly honest, perfectly cleaned data and still mislead — because the mapping from a real number to a stretched axis, a distorted 3-D angle, or an illegible overlay is a separate decision from whether the underlying data is trustworthy. Auditing a chart for honesty is a distinct skill from auditing the data behind it, and this exam tests both.

L1 — Intuition

Everything in sections 1 through 6 assumes a single tabular dataset with columns you can plot directly. A multimodal analysis — the kind this exam's Domain 4 material builds toward — usually needs to report on several different things at once: how well a model performs, how the training data is composed across modalities, and where attention or errors concentrate. The six-way mapping still applies to each individual question; what changes is that a multimodal report typically needs several of these chart types side by side, each answering a different piece of the same overall analysis.

L2 — Mechanism

Consider a status report on a text-to-image training run, which needs at minimum: the loss curve over training steps (a line chart — change over time), the distribution of caption lengths in the training corpus (a histogram — distribution of one variable), a breakdown of how many training examples come from each of several source datasets (a bar chart — compare categories), and a correlation matrix between several image-quality metrics being tracked (a heatmap — correlation across many variables). None of these four charts substitutes for another, and combining them into one chart — say, trying to show the loss curve and the caption-length histogram on the same axes — would violate the same question-to-chart mapping this lesson opened with, because a line chart's ordered x-axis (training step) and a histogram's binned x-axis (caption length) are answering two genuinely different questions and have no shared axis to plot on.

The multimodal-specific wrinkle worth naming is that a report combining modalities often needs one more chart type this lesson has not yet introduced by name: a direct visualization of where a model's attention concentrated on an image or a sequence of tokens, typically rendered as a heatmap overlaid on the original input rather than a heatmap of a correlation matrix. This is the same base chart type — color intensity encoding a value in a grid — reused for a different purpose: instead of encoding "how correlated are these two variables," the color intensity encodes "how much attention weight did the model place on this region." M2-04 covers this specific application, called an attention map, in full; the point worth carrying forward here is that "heatmap" as a chart type is broader than "correlation matrix" as a use case, and recognizing the shared visual grammar between the two helps you avoid treating them as two unrelated chart types when a scenario describes one or the other.

L3 — The exam-relevant edge case

A scenario describing a dashboard or report that mixes several charts is testing whether you can decompose it into its individual question-to-chart mappings rather than judging the report as one monolithic visual. If a described report's line chart correctly shows loss over time but its accompanying bar chart truncates its y-axis, the correct diagnosis names the specific truncated-axis problem in the specific bar chart — not a vague complaint about the whole report — because the exam's scenario wording is typically built around one specific, identifiable flaw per question, not a global critique.

Data Analysis and Visualization sits inside Domain 2 at 10% exam weight, and the source material frames chart selection as squarely foundational-level: "recognize the right technique and read a chart correctly," not build a custom visualization pipeline [GROUND TRUTH] (Sources/nca-genm/domain-2-data-analysis.md). The domain's own self-check question makes the expected question shape explicit: "A chart shows sales over 12 months. Which chart type fits best?" with the keyed answer "Line chart," against distractors offering a pie chart, a scatter plot, and a heatmap [GROUND TRUTH] (Sources/nca-genm/domain-2-data-analysis.md) — each distractor is a real, correct chart for some other question, misapplied here to a time-series question it does not fit.

The question tends to arrive in two shapes: a scenario describing an analytical goal in prose, asking which chart type to use (testing the six-way mapping directly), or a scenario describing a specific chart's construction — its axis range, its dimensionality, its density — and asking what is wrong with it or what conclusion it improperly supports (testing the misleading-visuals list). A less common but real third shape gives you a chart's description and a stated audience, testing the "simple charts suit general audiences" judgment independently of whether the chart is honestly drawn.

What the distractors typically look like

The reliable traps mirror the ones the source material calls out directly: offering a chart type that is correct for a related but different question (a scatter plot when the real ask is a trend over time, since both technically involve two axes); presenting a truncated-axis chart's exaggerated visual gap as if it reflects the real numeric gap; and describing an overloaded or 3-D chart's problems as being about the underlying data's quality, when the data itself may be perfectly clean and the failure is purely in how it was drawn.

A worthwhile habit for approaching any chart-selection stem quickly: identify the number of variables named in the scenario first, then the type of each. One continuous variable and a question about its shape or spread points to a histogram or box plot before you have even read the rest of the sentence. Two continuous variables and a question about whether they move together points to a scatter plot. One categorical variable paired with one numeric variable, and a question comparing groups, points to a bar chart. Any variable explicitly described as ordered by date, step, or session points to a line chart regardless of what else is in the sentence, because "how something changes as X increases" is the line chart's exact identity. This variable-counting habit resolves the large majority of six-way-mapping questions in a few seconds, leaving the harder judgment calls — stacked bar versus pie, histogram versus box plot for a distribution question that also mentions outliers — as the only ones that need the fuller reasoning in sections 1 and 2.

A box plot, not a histogram, is the standard answer, because several box plots line up side by side and remain individually readable, while overlaying several histograms on one axis quickly becomes an illegible stack of overlapping bars. The tradeoff is real: a box plot's five-number summary hides multi-modal shape (two peaks within one group's distribution are invisible in its box plot), so if the shape itself — not just the spread and outliers — is the specific thing being compared across groups, small multiples of separate histograms, one per group, is the more honest choice even though it takes more visual space.

Yes, for a small number of categories — commonly cited as two or three — where the part-to-whole relationship is the entire point and the audience needs an immediate, intuitive read rather than a precise numeric comparison. Past three or four slices, angle-based perception degrades quickly enough that a stacked bar chart, which encodes the identical part-to-whole data through length, becomes the more accurate choice, which is exactly why the source material's own guidance qualifies pie charts with "sparingly" rather than ruling them out entirely.

The distinguishing check is whether the visual gap matches the numeric gap once you read the axis carefully — a chart is not misleading merely because it shows something dramatic; it is misleading when the visual impression of the size of a difference does not match its numeric size. A bar chart with an honest zero-based axis showing one bar at 20 and another at 80 is showing a real, large, honestly-represented difference. The identical numbers, drawn with a y-axis running from 15 to 85, would compress that same real gap into a visually modest-looking difference — the reverse manipulation of the truncation example in section 6, and equally dishonest even though it understates rather than exaggerates.

A table of numbers requires the reader to hold every value in memory and mentally compute whether a pattern exists across dozens or hundreds of rows, which is exactly the kind of task human perception is bad at compared to reading position on a page. A scatter plot converts the same numbers into a spatial pattern — a tight diagonal band, a loose cloud, a curve, two separate clusters — that the visual system recognizes essentially instantly, and it does this without discarding any individual data point the way a single summary statistic like Pearson r does. The tradeoff appears only at very large row counts, where thousands of overlapping points can themselves become an overloaded chart in the sense section 3 describes, at which point a density-shaded heatmap-style scatter (sometimes called a hexbin plot) becomes the more legible choice — a specialized case, not a change to the core mapping.

Mostly yes, with one adjustment worth naming directly. The mapping in section 1 is written around a numeric variable on at least one axis — distribution and relationship both assume continuous values — but "compare categories" already covers the fully categorical case (one categorical axis, one numeric value per category), and a heatmap comfortably extends to two categorical axes with a numeric value in each cell, which is exactly the shape of a confusion matrix: rows are true classes, columns are predicted classes, and cell color encodes count. A confusion matrix is, mechanically, the identical chart type as a correlation heatmap — a grid with color-encoded cells — reused for a different pair of categorical axes, which is a useful example of how few genuinely distinct visual grammars this six-way mapping actually rests on once you notice the shared structure between "correlation across variables" and "agreement across predicted-versus-true categories."

• The six-way mapping is directly testable and worth memorizing cold: histogram/box plot (distribution), scatter (relationship), bar (categories), line (time), heatmap (correlation across many variables), stacked bar/pie sparingly (part-to-whole).
• A box plot's advantage over a histogram is comparability across groups; its cost is hiding multi-modal shape within any one group.
• A chart can come from honest, correctly cleaned data and still mislead — chart-type selection and honest scaling are two separate decisions, and this exam tests both independently.
• Truncated/zoomed axes exaggerate real differences by narrowing the visible range around the data rather than starting from a natural baseline.
• 3-D effects distort a chart's apparent values through the added depth dimension, compounding whatever perceptual disadvantage the base chart type (especially pie charts) already carries.
• Overloaded charts remain technically complete while becoming practically illegible — the fix is faceting or reducing scope, not a fancier single chart.
• Match chart complexity to the audience: simple charts for general audiences, more complex ones (like a large heatmap) only where the audience has the expertise to read them correctly.

You now know how to represent honestly cleaned, honestly summarized data as a picture that does not distort it. The next question in this module shifts from a picture of one modality to a picture of what a model is doing internally when it processes multimodal input — M2-04 covers attention maps, the heatmap-shaped tool used specifically to visualize where a transformer or vision-language model's weight went, and the specific way that visualization is suggestive evidence rather than a guaranteed causal explanation of the model's output.