Back to Course CMSC 178DA

Week 5: Data Visualization Principles

1 / 43
CMSC 178DA | Week 05

The Art of Insight

Strategic Principles of Data Visualization

Department of Computer Science

University of the Philippines Cebu

Visual Design & Perception

Why We Visualize

"The greatest value of a picture is when it forces us to notice what we never expected to see."

-- John Tukey

Agenda

Lecture Objectives

Perception

Pre-attentive attributes, Gestalt principles, color theory, and accessibility.

Design

Tufte's data-ink ratio, visual hierarchy, and minimalist frameworks.

Selection

Chart taxonomy, narrative structure, and strategic storytelling.

Moving from "making charts" to designing insights.

Two sessions covering professional visualization standards.

Part I

The Science of Seeing

Our brains evolved to detect patterns instantly -- but we must design for how humans actually see, not how we think they see.

This section covers pre-attentive processing, Gestalt principles, color theory, and accessibility.

I. Perception

Pre-attentive Processing

Pre-attentive attributes demo: color pop-out and size pop-out

Visual attributes like hue, size, and position are processed in <200ms -- before conscious attention begins.

The Insight

Effective visualization uses these attributes to highlight the "story" so the viewer perceives it instantly.

I. Perception

Gestalt Principles

Six Gestalt principles: proximity, similarity, enclosure, continuity, closure, connection

Humans naturally perceive holistic patterns. We group objects by proximity, similarity, and connection.

Pro Tip

Use whitespace (proximity) to group related data without heavy boxes or borders.

I. Perception

The Three Color Palettes

Three color palette types: sequential, diverging, categorical

1. Sequential

Low-to-high magnitudes. Use for population density, temperature, income.

2. Diverging

Extremes from a midpoint. Use for profit/loss, above/below average.

3. Categorical

Distinct unordered groups. Max 8-10 colors for readability.

I. Perception

Text is UI

The eye hates "ping-ponging" between a legend and the data.

Direct Labeling

Place labels directly next to lines or bars. Eliminate the legend entirely when possible.

Annotations

Don't just show "Sales Dropped." Add context: "Sales dropped due to supply chain outage."

Product A Product B Product A Product B
I. Perception

Inclusive Design

Colorblind simulation: normal vision vs deuteranopia comparison

8% of males have color vision deficiency. Designing with red-green scales excludes a significant portion of your audience.

TL;DR

Redundant Encoding

Never use color alone. Combine shape + color or pattern + color to ensure clarity for all viewers.

II. Frameworks

Tufte's Data-Ink Ratio

Data-ink ratio comparison: cluttered chart junk vs clean Tufte-style chart

"Above all else, show the data." Maximize the share of ink dedicated to actual data points over total ink used.

Erase

Gridlines, redundant labels, 3D effects, background gradients, and decorative elements.

II. Frameworks

The Active Title

Stop labeling charts with "What" (e.g., "Revenue by Quarter"). Label them with "So What?"

The Rule

If the title was the only thing someone read, would they get the point?

Q4 Revenue Breakdown (in Millions)

0 50 Retail Services Tech Other
VS

Revenue exceeded targets in Q4, driven by Tech sector

Q4 Target Retail Services 80M Tech Other
II. Frameworks

Visual Hierarchy

Guide the viewer's eye in a deliberate order using size, position, and color.

TL;DR

Reading Order

  1. Title -- What is this about?
  2. Main insight -- The key message (largest, boldest)
  3. Supporting detail -- Context and nuance
  4. Source / Notes -- Credibility (smallest)
1

Course Completion Rate

2
84%
3
+12% vs last semester
4

Source: Learning Management System database (Fall 2023). Excludes audited enrollments.

II. Frameworks

Focus Through Gray

Customer Acquisition Cost Spike in Q4

$0 $100 $200 Q1 Q2 Q3 $240 Q4 Q1 Q2

Use gray for context data and a bold accent color only for the data point you are discussing.

TL;DR

Technique

"Grey it out." If everything is colored, nothing is emphasized.

Knowledge Check

Tufte's Data-Ink Ratio

Which of the following elements is considered "Chart Junk" and should be erased to maximize the data-ink ratio?

A) Data Line
Y X B) Axis Labels
C) Background Gradient
+20% D) Trend Annotation
Click & hold to reveal answer

Correct Answer: C

Background gradients don't represent data. They are purely decorative "ink" that distracts from the actual information.

Part III

Choosing the Right Chart

Chart selection is strategic, not aesthetic. The wrong chart can mislead even with perfectly correct data.

From bar charts to treemaps -- every form serves a distinct analytical purpose.

BAD 3D Exploding Pie (Hard to read)
GOOD Clean Bar Chart (Easy to compare)
III. Selection

Chart Taxonomy & Questions

Goal Chart Types Primary Question Answered
Comparison Bar, Grouped Bar, Lollipop "Which [category] is the highest/lowest?"
Distribution Histogram, Box, Violin "How is the data shaped or spread?"
Relationship Scatter, Bubble, Heatmap "Does A correlate with B?"
Composition Stacked Bar, Treemap, Donut "What makes up the total?"
Trend Line, Area, Sparkline "How does this change over time?"

Start with the question, then pick the chart -- never the other way around. Form follows function.

Quick Test

Ask: "What question does this chart answer?" If you can't answer clearly, pick a different chart.

III. Selection

Chart Type Gallery

Six chart types: bar, histogram, scatter, box, line, heatmap

All examples use Philippine economic data from the Philippine Statistics Authority (PSA 2023).

Notice

Each chart type serves a distinct analytical purpose -- they are not interchangeable.

III. Selection

Precise Comparisons

Horizontal bar chart of Philippine GDP by region

Bar charts are the gold standard for comparing discrete values. The human eye is highly accurate at comparing lengths.

Golden Rule: Always start the Y-axis at zero for bar charts.

III. Selection

Distributions & Spread

Distribution charts: histogram, box plot, violin, KDE

Histograms and KDE plots show the "shape" and modality of data. Is it Normal, Skewed, or Bimodal?

Why It Matters

Distributions dictate your statistical strategy. Skewed data needs non-parametric tests.

III. Selection

Scatter Plots & Correlation

Scatter plot with trend line and correlation heatmap

Scatter plots reveal relationships between numeric pairs. Use hue and size to add dimensionality.

Insight

Clusters and outliers often tell a more interesting story than the trendline itself.

III. Selection

Composition & Trends

Stacked bar chart and line chart showing composition and trends

Stacked area and bar charts reveal how parts of a whole change across categories or over time.

Limitation

Middle categories lack a common baseline, making them hard to compare. Use with caution.

III. Selection

The Pie Chart Problem

The human brain is terrible at comparing angles and areas.

Avoid Pie Charts If:

  • You have more than 3 slices
  • Values are similar (e.g., 24% vs 26%)
  • You want to compare precise values

Better Alternative

Horizontal bar charts or donut charts with direct labels.

III. Selection

Small Multiples (Faceting)

FacetGrid: Philippine family income distribution by island group

Small multiples repeat the same chart across categorical slices, enabling high-dimensional comparison.

Efficiency

Consistent axes let the eye scan for subtle differences across groups without overlap.

III. Selection

Heatmaps & Tables

Sometimes you need the raw numbers and the visual pattern.

Visual Table

Apply a sequential color scale to cell backgrounds to highlight high/low values instantly.

12
35
62
84
18
42
71
94
22
48
76
105
IV. Integrity

Common Mistakes to Avoid

Truncated Y-Axis

Exaggerates small differences, misleads viewers about magnitude

Dual Y-Axes

Creates false correlations between unrelated data series

3D Charts

Distorts perception, occludes data points, adds zero information

Pie Chart Overload

Too many slices making it impossible to compare angles and areas

Rainbow Scales

Unintuitive, lacking natural order, and often not colorblind-safe

Overcomplicated

Cramming too much information without a clear visual hierarchy

The most common visualization errors fall into two categories: distorting reality (lying with data) and cognitive overload.

Truth over impact. Maintain proportional honesty in all reporting.

IV. Integrity

The Truncated Axis Trick

Side-by-side: misleading truncated y-axis vs honest full axis

Truncating the Y-axis is the most common visual lie. It exaggerates small differences into false "massive" trends.

The Same Data

Same numbers, dramatically different impression. Context and axis range change everything.

IV. Integrity

The Dual Axis Trap

Misleading (Dual Axis) Ice Cream Sales Shark Attacks 0 100k 0 10

Forces lines to cross or align, implying false causation.

Honest (Two Charts)
Sales Attacks

Allows viewers to see trends independently.

Two completely unrelated variables can be made to look perfectly correlated simply by adjusting the scaling of dual Y-axes.

IV. Integrity

The 3D Distortion Effect

Distorted (3D)
A B

Slice A physically takes up 70% of the ink, but represents only 30% of the data.

Accurate (2D Flat)
A B

Honest representation where visual area cleanly maps to data area.

3D rendering tilts the chart, pulling the front elements physically closer to the viewer. This violates the principle of proportional honesty.

IV. Integrity

Pie Chart Overload

Cognitive Overload
Item A
Item B
Item C
Item D
Item E
Item F

Humans cannot accurately compare angles of similar sizes, forcing eyes to dart back and forth to the legend.

Ranked Clarity
Item A
Item H
Item C
Item B
Item F
Item D
Others

Direct translation of length allows instant, effortless ranking. Text labels are placed right beside the data.

Rule of thumb: If a pie chart has more than 4 or 5 slices, use a bar chart instead. Never make your audience do the visual math.

IV. Integrity

Simpson's Paradox

A trend appears in groups of data but disappears or reverses when groups are combined.

The Lesson

Aggregated data can lie. Always check for confounding variables that might skew the result.

Aggregated: Trend is POSITIVE
Disaggregated: Trends are NEGATIVE
Knowledge Check

Choosing the Right Chart

You need to show the correlation between study hours and final grades for 500 students. Which chart type should you use?

A) Bar Chart
B) Pie Chart
C) Scatter Plot
D) Line Chart
Click & hold to reveal answer

Correct Answer: C

A scatter plot is the standard choice for showing correlation and relationships between two continuous numeric variables.

V. Advanced

Embracing Uncertainty

Data is rarely precise. Don't hide the variance.

Error Bars

Use 95% Confidence Intervals to show the range of the true mean, not just the sample mean.

Warning: Standard Deviation (SD) and Standard Error (SE) are not the same. Know what you are plotting.

High Variance Low Variance
V. Advanced

Visualizing Flow

Input Output A Output B

Sankey Diagrams quantify the flow of resources (budget, energy, users) through a system.

Use Case

"Where did the budget go?" or "Where are users dropping off?" Width = magnitude of flow.

V. Advanced

Hierarchical Data

Pie charts fail at showing nested data with many categories. Treemaps excel at it, like showing the breakdown of a country's exports.

TL;DR

Why Treemaps?

They use 100% of the space to show part-to-whole relationships, scaling easily to dozens of categories (like the Philippine export economy).

Philippine Exports (2023)
Electrical & Electronics 56%
Machinery 9%
Metals/Ores 6%
Agri & Food 11%
Other Manufactures 18%
V. Advanced

Ranking Many Categories

With 10+ categories, bar charts can feel visually heavy. Lollipop charts reduce visual weight while emphasizing the value.

Design

The dot (value) becomes the focal point while the stem provides alignment. Less ink, same information.

V. Advanced

Spatial Data

Maps are powerful but dangerous. Large geographic areas dominate visually despite low data importance.

The Choropleth

Encodes data values into geographic regions using color intensity. Works best for density-based metrics (per capita, rates).

Geographic Context Interactive maps display true area, exposing how massive regions can dominate visual weight regardless of underlying data density.
Part VI

Driving Action with Narrative

A chart is a character; context is the plot. Without a narrative arc, data points remain static observations instead of catalysts for action.

Structure your presentations like a story -- context, conflict, resolution.

DATA DUMP (Boring)
THE STORY (Actionable) Sales Spiked!
VI. Narrative

Data Storytelling

Charts are not the destination; they are the vehicle for the insight.

The 3-Act Structure

  1. Context: What is the baseline?
  2. Conflict: What changed? (The anomaly)
  3. Resolution: What do we do? (The recommendation)
1. Context The Baseline 2. Conflict The Anomaly 3. Resolution The Action
VI. Narrative

Annotation Best Practices

0 4M 8M 2018 2019 2020 2021 2022 International Tourist Arrivals (PH) 82% Drop COVID-19 Global Travel Restrictions

Annotations transform a chart from "here's data" to "here's what the data means."

Techniques

Arrows for key events, shaded regions for periods, direct labels for interpretation.

VII. Application

Strategic Dashboards

Philippine tourism analytics dashboard with KPIs and charts

A dashboard should answer the most critical KPI question in under 5 seconds. Hierarchy is everything.

Hierarchy

Big number (most critical metric) in the top-left -- the natural focal point of attention.

VII. Application

Anatomy of a Dashboard

Tourist Arrivals
5.45M
+15.1% YoY
Revenue
P271B
+22.3% YoY
Occupancy Rate
68.4%
-2.1pp vs target
Avg Stay
4.2 days
+0.3 days
Monthly Arrivals Trend (2023)
Top Destinations
NCR
42%
Cebu
28%
Boracay
18%
Palawan
12%

KPI cards + trend line + breakdown bars -- every panel answers one specific question.

Structure

Summary metrics at top, trends in middle, detail breakdowns at bottom.

VII. Application

Plotly: Interactive by Default

For web presentations, static images (like Matplotlib PNGs) restrict the audience. Modern declarative tools build interactive charts natively in the browser.

Built-in Capabilities

Hover tooltips, click-and-drag zooming, panning, and responsive scaling come for free without writing extra interactive logic.

VII. Application

Seaborn: Statistical Visualization

Seaborn regplot with regression line and boxplot by island group

Seaborn automatically adds regression lines, confidence intervals, and statistical groupings.

When to Use

Statistical plots (regplot, boxplot, violinplot) where built-in analytics add value beyond matplotlib.

VII. Application

Interactive Visualization

1. Overview

See the full picture first

2. Zoom & Filter

Narrow to areas of interest

3. Details

Inspect specific data on demand

Shneiderman's Mantra: "Overview first, zoom and filter, then details-on-demand."

Tools

Plotly Express for hover tooltips, zoom, pan, and export. Dash or Streamlit for full interactive dashboards.

VII. Application

Philippine Example: Regional GDP

Gross Regional Domestic Product (2023)
Region Share (%)

Sorted horizontal bars with direct labels -- no gridlines needed. NCR dominates at over 5x the next region.

Source

Philippine Statistics Authority (PSA), 2023 Regional Accounts. Design: sorted, labeled, colorblind-safe.

Activity

Fix This Chart

Imagine a 3D pie chart with 12 slices, rainbow colors, no labels, and the title: "Data".

  1. List at least 3 problems with this chart.
  2. How would you redesign it?
  3. Which chart type would be a better choice?
3 minutes

Apply everything we covered: data-ink ratio, color choice, chart selection, and active titles.

Discuss

Share your redesign with a neighbor. Did you pick the same chart type?

Industry Tools

The Modern BI Ecosystem

Why leave Python/R?

Libraries like Matplotlib, Seaborn, and Plotly are excellent for ad-hoc analysis and data science workflows. However, deploying insights to hundreds of non-technical stakeholders requires a different toolset.

Scale

Cloud-native architecture to handle millions of rows without crashing the user's browser.

Governance

Centralized semantic models ensure everyone agrees on the definition of "Revenue" or "Active Users".

Interactivity

Self-service drag-and-drop filtering, cross-filtering, and exporting for business users.

BI (Business Intelligence) platforms are the enterprise standard for operational reporting.

Industry Tools

Tableau: The Artist's Choice

Tableau revolutionized the BI industry by turning drag-and-drop actions into database queries.

Strengths

Unmatched visual flexibility. The famous "Show Me" feature and its underlying proprietary VizQL (Visual Query Language) allow for rapid, intuitive exploratory data analysis.

Trivia

Founded in 2003 by researchers from Stanford University who commercialized their Department of Defense-funded research. Acquired by Salesforce in 2019 for $15.7 Billion.

Tableau Sample Dashboard

Tableau

"Help people see and understand data."

Power BI Sample Dashboard

Power BI

The Enterprise Heavyweight

Industry Tools

Power BI: The Corporate Standard

Microsoft's flagship analytics tool tightly integrated with the Azure and Office 365 ecosystem.

Strengths

Incredibly powerful data modeling via DAX (Data Analysis Expressions) and Power Query (M language). Highly cost-effective for enterprises already using Microsoft E5 licenses.

Trivia

Originally designed under the code name "Project Crescent". It evolved into a behemoth by combining several obscure Excel add-ins (Power Pivot, Power Query, Power View).

Industry Tools

Looker: The Developer's BI

A 100% web-based platform that forced the industry to rethink how core metrics are governed.

Strengths

Introduced LookML, a Git-version-controlled semantic layer. This ensures a metric like "Revenue" is defined exactly once in code, meaning every dashboard across the company calculates it exactly the same way.

Trivia

Unlike Tableau or Power BI, Looker doesn't extract data into its own memory engine—it translates LookML into SQL and queries the data warehouse (like BigQuery or Snowflake) directly. Google Cloud acquired Looker in 2019 for $2.6 Billion.

Looker Sample Dashboard

Looker

Governance First, Viz Second

Key Takeaways

  1. Pre-attentive attributes enable instant perception (<200ms)
  2. Tufte's data-ink ratio: maximize data, minimize junk
  3. Match chart type to your message and data type
  4. Use active titles and annotations to tell the story
  5. Always design for accessibility and integrity

Lab 5: Visualization Portfolio

Create publication-quality charts with matplotlib & seaborn using Philippine regional data. Practice chart selection, Tufte's principles, and data storytelling.

tufte.com

seaborn.pydata.org

plotly.com