Post

DAVI Visualizing Time-oriented Data Redo

Posted Updated
By Wei Xiong
21 min read
DAVI Visualizing Time-oriented Data Redo

Lecture Notes: Time Visualization and Visualization Critiques

Course: DAVI (Data Visualization)
Professor: (“the professor” as referred to in the transcript)
Date: (Exact date not specified in the transcript)

Below are detailed notes capturing everything the professor covered, in chronological order, along with relevant sketches, explanations, and critiques. These notes include:

  • Recap questions and their answers
  • Discussions on various charts (parallel coordinates, mosaic plots, marimekko charts, trellis plots, etc.)
  • Detailed exploration of time-oriented data visualization (including tasks, data characterization, and multiple examples)
  • Interactive techniques (e.g., stack zooming, ChronoLenses)
  • Two visualization critiques (analyzing faulty graphs)

Use these notes to thoroughly understand the concepts, examples, and tips mentioned by the professor.

0:00 – 0:49: Introduction and Weekly Plan

  • Greeting: The professor says, “All right. Good morning. Good morning everyone.”
  • Plan for the week:
    • Today:
      • Time visualization (visualizing temporal data).
      • Two visualization critiques (in the third hour).
    • Wednesday:
      • Design exercises in preparation for an industry customer visiting next week.
    • Emphasis on “getting creativity engaged and activated” before the industry guest arrives.
  • Lecture Recap: The professor will do a short lecture recap using Mentimeter (“mentees”).
  • The code is at the top, or students can use the Brightspace link to join.

0:50 – 3:05: Recap Question 1

Question: In a parallel coordinates (perl coordinates) plot, the shown pattern indicates:

  • A cluster
  • A positive correlation
  • A negative correlation
  • A high variance of spread
  • Answer: Negative correlation.
  • Explanation:
    • In parallel coordinates, if two attributes are negatively correlated, high values on one axis align with low values on the other axis and vice versa, forming a star-shaped crossing pattern.
    • A positive correlation would look more like a “rather horizontal line” (attributes not crossing).
    • High variance of spread refers to a single variable’s variance, not the relationship between two variables.
    • It is not a cluster, because that pattern specifically indicates negative correlation.

3:06 – 4:07: Recap Question 2

Question: “What chart is this?” (Shows a screenshot with subdivided widths and heights)

  • Stacked Bar Chart
  • Stacked Column Chart
  • Mosaic Plot
  • Marimekko Chart
  • Answer: Marimekko chart (also a “flavor” of a mosaic plot).
  • Key Points:
    • Mosaic plots are for categorical data (splitting categories along x and y to subdivide rectangles).
    • Marimekko charts apply a similar principle to quantitative data by binning the data range into intervals (like quartiles or equal bins). Each bin acts as a category subdivision.
    • Compare to stacked bar/column charts:
      • In stacked charts, all columns have the same width. A Marimekko or mosaic plot adjusts widths (and possibly heights) based on proportions, so widths can vary.
    • Online, “mosaic plot” and “marimekko chart” are often used interchangeably.
    • The big clue is the variable widths and heights subdividing the space.

4:08 – 5:37: More on Marimekko vs. Mosaic

  • Why not “stacked bar chart”?
    • Because stacked bar charts keep a uniform width (or height), while mosaic/Marimekko subdivide horizontally and vertically in proportion to the data.
  • Remember:
    • Mosaic = categorical.
    • Marimekko = binned quantitative data → effectively “categorical” after binning.

5:38 – 6:06: Recap Question 3

Question: Parallel sets are used for:

  • Text data
  • Discrete data
  • Continuous data
  • Qualitative data
  • Answer: Qualitative (categorical) data.
  • Explanation:
    • Parallel sets are an alternative to mosaic plots for multiple categorical variables. Example used: The Titanic dataset (male/female, survived/perished, passenger class, crew, adult/child, etc.).
    • You can use parallel sets for numeric data only if you bin the numeric values into categories/intervals (e.g., age grouped into brackets).
    • They originally target categorical data.

6:07 – 7:07: Recap Question 4

Question: The trellis plot:

  • (A) Benefits from brushing and linking
  • (B) Subdivides the dataset
  • (C) Is mirrored around the main diagonal
  • (D) Works only for categorical data
  • Answer: (B) It subdivides the dataset.
  • Details:
    • Trellis plots group subsets of data into small multiples laid out in a grid (e.g., subdividing by categories or binned continuous variables).
    • They do not benefit from brushing and linking in the same way as scatterplot matrices do, because each small subplot has its own unique subset (no item belongs to multiple panels).
    • Mirroring around the main diagonal is for scatterplot matrices, not trellis plots.
    • Trellis plots are not limited to strictly categorical data; continuous data can be binned to create “trellis panels” as well.

7:08 – 8:00: Recap Question 5

Question: In Fun(B)ike’s model of visualization, which aspect of the visualization pipeline is detailed?

  • The mapping step
  • The interaction
  • The use of the visualization
  • The “understand” step
  • Answer: The use of the visualization.
  • Explanation:
    • The standard visualization pipeline (raw data → transformations → mapping → rendering → view) is quite condensed about what the user actually does.
    • Fun(B)ike’s model expands on how users perceive the image, generate knowledge, explore, refine specifications, etc. It blows up the “user” part of the pipeline.

8:01 – 9:15: Examples and Clarifications

  • The professor revisits the difference between the basic pipeline (data → visualization → user) and Fun(B)ike’s approach, which highlights knowledge generation, perception, exploration as separate steps.
  • One note: The “understand” step is from the “design activity framework,” not from the standard pipeline or from Fun(B)ike specifically.

9:16 – 9:52: End of Recap, Transition

  • Top quiz performers: Mentions “Eleazar” among top quiz scorers (just a casual mention).
  • Moves on to the main topic: Time Visualization.
  • Clarifies that from now on, each lecture condenses an entire research field (e.g., time-oriented data, geospatial data) into roughly 90 minutes.

9:53 – 10:40: Reference Books on Time Visualization

  • Shows two books:
    1. Visualizing Time
    2. Visualization of Time-Oriented Data (by Wolfgang Aigner, Silvia Miksch, Heidrun Schumann, Christian Tominski).
  • Mentions that it is from Springer, free via university network; one is completely open access.
  • Recommends Visualization of Time-Oriented Data because two authors overlap with the main course textbook, ensuring consistent coverage.

10:41 – 11:33: Why Time Visualization?

  • Many real-world data sets have a temporal dimension.
  • We see many bad time visualizations. Need to know how to do it properly.
  • The professor references an “egregious” example from a previous lecture to underscore the importance of learning correct techniques.

11:34 – 12:00: Outline for Today’s Time Visualization

  1. Principles of time visualization (data and task characterization).
  2. Time intervals visualization (specific techniques).
  3. Event-based data (briefly).
  4. Interaction techniques for time-oriented data (briefly).

12:01 – 13:02: Principles: “What is Time?”

  • Philosophical question, but from a data perspective we consider timescales:
    • Ordinal: We only know the order (before/after) but not exact intervals.
    • Discrete: Distinct time steps with measurable intervals (but data might only exist at those steps).
    • Continuous: Time flows with possible interpolation (e.g., measuring temp across the day).
  • We typically do not have a purely categorical scale for time, because time intrinsically has an order.

13:03 – 14:09: Time Scope

  • Distinguishes between time points vs. time intervals:
    • A reading or activity that’s valid for one “instant” (e.g., a measurement at midnight).
    • A reading or activity valid over a range (e.g., drilling a hole from 10:00–10:15).
  • Sometimes measurements are taken discretely, but we treat them as intervals valid until the next measurement arrives.

14:10 – 15:00: Time Arrangement

  • Linear: Our usual timeline, continuous from past to future.
  • Cyclic: Time repeats in cycles (days, seasons, fiscal years, etc.).
  • Branching: Especially in version control (Git), we have forks, merges, separate branches of time.

15:01 – 16:09: Time Granularity

  • Time can be measured by days, weeks, months, years, etc.
  • A calendar is effectively multiple granularities.
  • Important because tasks might require seeing data at multiple scales (daily/weekly/yearly).

16:10 – 16:53: Data Examples

  • Example:
    • Left chart: Continuous time scale, time points (scope) → typically a line chart with dots at measurement times.
    • Right chart: Discrete time scale, intervals (scope) → bar steps that remain constant until the next measurement.
  • Question: Can we combine a continuous time scale with an interval scope?
    • Yes, e.g., a Gantt chart: intervals that start anywhere on a continuous axis and last for variable lengths (like daily runs at random times).

16:54 – 17:58: Gantt Chart Example

  • A Gantt chart is a prime example of:
    • Continuous timeline.
    • Activities with intervals that can start/end at arbitrary times.
  • So it’s fully possible to have continuous + intervals.

17:59 – 20:03: Tasks with Temporal Data

  • All standard tasks apply (finding extremes, reading values, etc.), plus:
    • Synoptic tasks: Observing behavior over time (changes, trends).
  • Two general approaches:
    1. Fluctuating data → Searching for patterns or “homogeneity” (trends, periodicities, sequences).
    2. Homogeneous data → Searching for “heterogeneity” or divergences (outliers, irregularities).

Homogeneity vs. Heterogeneity

  • Homogeneous scenario: e.g., daily temperature is very predictable. Then a sudden deviant day or deviant pattern is what stands out.
  • Heterogeneous scenario: data is all over the place. You look for stable repeating motifs or trends within the noise.

20:04 – 22:08: Examples of Tasks

  1. Periodicities: Repeating cycles (seasonal, daily, etc.).
    • Spiral visualizations (like Tominski’s interactive example) to visually align cycles by adjusting the spiral’s period.
  2. Sequences: Certain sequences of events that might signal something (like predictive maintenance in industrial settings).
  3. Outliers: Data that breaks a normal pattern (e.g., one day the temperature doesn’t follow day-night cycle).
  4. Deviation from frequency: “Jitter” in something that’s normally periodic.

22:09 – 23:33: Spiral Visualization Example

  • Showed the “Spiral Chart” for daily doctor visits:
    • When the spiral turn is set to 4 weeks, we see a strong Monday spike repeating each cycle.
    • By adjusting the spiral’s period, we can discover/verify different cyclical patterns.

23:34 – 24:39: Arc Diagrams Example (Sequences)

  • Arc Diagrams: Another technique to find repeating sequences or patterns over time.
  • Used with:
    • Text or music (repeated themes).
    • Precipitation data (days with similar rainfall).
  • Large arcs can show that a pattern from Day X reoccurs at Day Y.
  • Great for discovering repeated motifs in otherwise messy data.

24:40 – 25:06: Summarizing Tasks → Visualization

  • The specific tasks for time data (periodicity, sequences, outliers, etc.) guide which visualization technique is best.
  • Design depends on data + task.

25:07 – 27:30: Examples of Time-Oriented Visualizations

  • Website: timeviz.net (associated with the “Visualization of Time-Oriented Data” book).
  • Contains a large gallery of time visualization techniques:
    • Arc diagrams, spirals, timeline-based charts, etc.
  • End of the first half: “After the break, we’ll look at how to map time to visual attributes.”

27:31 – 29:21: Mapping Time to Visual Attributes (Post-Break Start)

  • Revisits Mackinlay’s ranking of visual channels:
    • Position (x or y) is the most common for time (like a timeline).
    • Length is meaningful when dealing with durations.
    • Angle can represent time, e.g., clocks or radial cycle charts (like Apple Watch “activity rings”).
    • Connection: connected scatter plots (time as an implicit order, connecting data points in their chronological sequence).
    • Size, color: possible, though color for time is typically less effective unless everything else is “taken.”

29:22 – 31:03: Don’t Use Shape for Time

  • Shape is an identity channel, not an ordered magnitude channel.
  • Time is inherently sequential/ordered, so shape is ill-suited to convey before/after or more/less.

31:04 – 34:37: Time → Color Example

  • Example: “Travel time map” to a city center.
    • Each pixel is colored by how long it takes to travel to the nearest city (≄ 50,000 population).
    • Here, we use color because position is already used by geography. Size also is not an option for single pixels.
  • Usually, we avoid color for time since it’s not as precise for showing magnitude or direct comparison—unless no other channel is feasible.

34:38 – 36:08: Time to Display Space (Small Multiples) vs. Time to Display Time (Animation)

  • Option 1: Mapping time to space → Small multiples:
    • Each time point becomes its own chart in a grid or sequence.
    • Easy side-by-side comparisons but can become very small as the number of time points grows.
  • Option 2: Mapping time to “display time” → Animation:
    • One chart that changes over time. Good at showing transitions, but harder to compare different frames directly.
  • They are not strictly exclusive. Rare example:
    • A combined approach, morphing from one small multiple to the next in an animated sequence.

36:09 – 38:01: Gapminder Example (Animation + Connected Scatter Plot)

  • Hans Rosling’s Gapminder:
    • Each bubble = a country.
    • Income (x-axis), Life expectancy (y-axis).
    • Animation from year to year.
    • Selected countries show a connected trajectory (connected scatter plot).
    • The rest remain just animated bubbles.
    • This approach merges “history” (the line for selected countries) with ongoing animation (all countries shifting each year).

38:02 – 40:03: Line Charts for Time Series

  • Line chart is extremely common for continuous time + continuous measurements.
  • Problem: If you have many time series (e.g., 50, 100, 500), a standard line chart can get cluttered.

Horizon Graphs

  • A space-efficient approach for many time series.
  • How it works:
    1. Convert your series to an area chart with color encoding for positive/negative ranges.
    2. “Fold” or “stack” the chart by layering intervals on top of each other, flipping negatives upward in color-coded form.
    3. Compress the vertical space drastically.
  • Often you also sort these horizon graphs by similarity to see groupings.

40:04 – 42:00: Calendar-Based Visualization (Multiple Granularities)

  • Example: A monthly calendar, each day having a mini time series (like an area for each hour).
    • Additionally, a daily average line in red.
    • So you see day-level aggregates + within-day trends.
    • Possibly weekly summaries, too.
    • Let’s you see patterns at daily or weekly or monthly scale.

42:01 – 46:06: Visualizing Time Intervals

  • Gantt Charts:
    • Common when there are intervals, each with start and end times, displayed as horizontal bars.
    • Works for 12–24 intervals but can break if you have hundreds or thousands.

Triangular Model

  • A technique to show many intervals as points instead of bars.
  • Coordinates:
    • x-axis = start time
    • y-axis = duration (end time − start time)
  • So each interval is just a single point in (start, duration) space.
  • You can reconstruct:
    • If intervals overlap, whether they contain each other, etc., by referencing diagonal regions in the triangular space.
  • Great for large sets of intervals; you can also do binning, color coding, heatmaps for dense overlap.

Reading the Triangular Model

  • If two intervals are the same, they map to the same point.
  • Intervals that “meet exactly” (one ends where the other starts) form patterns along diagonals.
  • Containing intervals = points above/below one another in this coordinate system.
  • Often accompanied by interaction to highlight relevant sets of intervals.

46:07 – 48:08: Triangular Model for Continuous Data (Multi-Resolution)

  • Example: Monitoring a soccer player’s speed second by second.
  • They aggregate smaller intervals into bigger intervals in a pyramid-like structure (1s, 2s, 4s, 8s
).
  • Color indicates average speed. The top row = entire match average, bottom row = second-by-second details.
  • Called “multiresolution time interval visualization,” effectively layering multiple granularities in a triangular layout.

48:09 – 49:21: Event-Based Data

  • Another category is event-based (like Twitter posts, error logs, etc.).
  • Two main examples:
    1. (Skipped in detail by the professor)
    2. Time Maps: A scatter plot approach for events.

Time Maps

  • Coordinates:
    • x = time since previous event
    • y = time until next event
  • If an account tweets repeatedly in quick succession, those points gather near the origin on one axis, etc.
  • Example: Barack Obama’s Twitter feed
    • Patterns: “First tweet of the day,” “Last tweet of the day,” bursts of frequent tweeting (clusters near small intervals).
  • Finding Twitter Bots:
    • A bot’s pattern can form a suspiciously regular shape in the time map, revealing non-human behavior.
    • No need for text analysis or follower counts—just the time stamps can suffice to identify bots.

49:22 – 51:16: Interaction Techniques for Time Data

  • Multi-level (Stack) Zoom (Nicholas Empringham approach):
    • A top-level timeline: user draws a region of interest.
    • Below, that region is “blown up” in a new sub-panel.
    • You can do multiple selections, each spawns another sub-panel, etc.
    • Great for retaining a high-level overview and multiple drilled-in views simultaneously.

51:17 – 55:31: ChronoLenses Demonstration

  • ChronoLens is a powerful interactive approach:
    • Lenses are overlaid on time series for magnification, filtering, or transformations (like cross-correlation).
    • You can chain lenses (the result of one lens becomes the input of another).
    • Each lens can be zoomed, moved, resized, or locked in place.
    • Users can build an analysis pipeline visually:
      • E.g., parent lens outputs a smoothed series → child lens cross-correlates with another series, and so on.
    • Very advanced for exploratory tasks (searching for hidden patterns or relationships).

55:32 – 56:22: Summary of Time Visualization

  • If you already know exactly what you want (a single metric to optimize), you may not need fancy visualization. But if you need to explore, see trends, or discover anomalies, interactive time visualization tools like ChronoLens are extremely valuable.

56:23 – 57:10: Next Week’s Topic

  • Geospatial data (map-based visualizations).
  • The course text has a short subchapter; recommended reading before class.

57:11 – 58:07: Literature References

  • The professor shows references from the slides:
    • “Multi-Dimensional Data Visualization” (book),
    • “Visualizing Time,”
    • “Visualization of Time-Oriented Data” by Aigner et al.,
    • And presumably the standard textbook by Munzner or Tamara Munzner’s “Visualization Analysis and Design” (not mentioned by name here, but often used).

58:08 – 59:04: Transition to Two Visualization Critiques

  • The class takes a short break and returns at 11:20 for critiques.

59:05 – 1:00:00: Visualization Critique #1: “America’s Religious Landscape”

  • The professor displays a chart that was published with the message:

    “No religious group is larger than those who are unaffiliated from religion.”

  • Students get 3 minutes to examine individually, 3 minutes to discuss with neighbors.

1:00:01 – 1:02:00: Identifying the Biggest Problem

  • Professor asks, “What’s the single biggest problem with this chart that sinks the ship?”

Chart Description

  • It’s a pie chart with multiple slices, each subdivided by ethnicity and religion in a single slice (e.g., “White Catholic,” “Black Protestant,” “Hispanic Protestant,” “White Evangelical,” etc.).
  • The final major slice is “Unaffiliated,” presented as one big portion.

Observed Problems

  1. Mixing religion and ethnicity in the same “top-level slices,” artificially inflating or splitting certain groups (Protestants or Catholics) by race, while “Unaffiliated” is not subdivided.
  2. The numeric totals sum to 102% (rounding error), which they didn’t even correct by going to decimal places.
  3. Color usage might be questionable in terms of colorblindness or ordering.

1:02:01 – 1:03:22: Structured Critique

  • Chart Type: Pie Chart.
  • Data: Grouped categories + percentages.
  • Task: Comparison overview, specifically to see if “unaffiliated” is the largest single group.
  • Expressiveness:
    • The biggest offense: They combined ‘White Catholic, Hispanic Catholic, Black Protestant, etc.’ while “Unaffiliated” is not broken down by race. So it’s inconsistent.
    • If they wanted a breakdown by race, do so for every group (including unaffiliated). If they only want a breakdown by religion, separate them consistently.
    • The purposeful mismatch likely artificially drives home the point that “unaffiliated” is the largest group.
  • Color Issues:
    • Possibly not colorblind-safe.
    • They used color “hue + brightness” in a way that might imply an ordering that doesn’t exist.
  • Rounding:
    • They show integer percentages that sum to 102%. Easy fix: show one decimal or correct the rounding.
  • Better Alternatives:
    • A Sunburst or Multilevel Donut: put religion in the inner circle, subdivide ethnicities in the outer ring if it’s truly needed.
    • Or do separate comparisons: one chart purely for religion, another purely for race.

1:03:23 – 1:05:30: Professor Checking the Source

  • The professor attempts to open the original PDF to confirm sample sizes, methodology.
  • Possibly thousands of participants, not just 102 people. “102%” is clearly a rounding artifact.
  • Concludes: The chart is from a broader study, but the figure caption does not clarify. The discrepancy is unaddressed.

1:05:31 – 1:07:00: Visualization Critique #2: Bivariate Map of Wisconsin

  • Second example: A map of Wisconsin’s counties (choropleth for median household income) with overlaid circles for “number of residents with a high school degree or equivalent.”

1:07:01 – 1:09:02: Big Problems

  1. No normalization: Circles show the absolute number of high school graduates. This mostly reflects population size, not the proportion of graduates.
  2. Circle size mapped to diameter instead of area, causing a “lie factor.”
  3. Binning for household income is questionable or inconsistent. Possibly not quintiles, not equal intervals—unclear logic.

1:09:03 – 1:11:20: Detailed Critique

  • Chart Type:
    • Choropleth map + overlaid circles (one numeric variable in color, another in circle size).
  • Data:
    • Median household income (quantitative).
    • High school graduates (quantitative).
  • Task:
    • Possibly correlation between education level and income, or to see distributions across the state.
  • Expressiveness:
    • Using absolute numbers misleads the viewer into reading large circles just because of big cities (Milwaukee, Madison, Green Bay).
    • The circle scaling is done by diameter (not area), which visually overemphasizes differences.
  • Effectiveness:
    • The color binning for household income is unclear; intervals are uneven, and it doesn’t look like standard quintiles or a known method.
    • Circle placement is also questionable; some circles overlap county borders unnecessarily.
    • For clarity, should have used proportion (percentage of population with HS degrees), and scaled circles by area.
  • The professor references how a better approach might be:
    • Use recognized binning strategies (equal intervals, quantiles, or standard breaks used by major statistical agencies).
    • Possibly partial displacement or an “expanded map” approach so circles don’t overlap boundaries, or use interactive labeling.

1:11:21 – 1:12:30: Summary and Closing

  • The professor provides links:
    • The map example source.
    • Info on quintile methods.

  • Encourages students to attempt to compute the “lie factor” for that second chart (due to diameter scaling) as an exercise.

  • Class ends with: “Have a nice week. See you on Wednesday for design exercises and the life-factor discussion.”

Key Takeaways

  1. Time Visualization:
    • Carefully define time scale (ordinal, discrete, continuous), time scope (point vs. interval), time arrangement (linear, cyclic, branching), and granularity.
    • Match tasks (finding trends, periodicity, outliers, etc.) to appropriate techniques (spirals, arc diagrams, calendar visualizations, triangular model, Gantt charts, time maps).
    • Consider advanced interaction (stack zoom, ChronoLens) to explore complex time series or event data.
  2. Common Pitfalls:
    • Using color for time can be tricky—do it only if position, size, etc. are unavailable or you have a strong reason.
    • Mapping time to shape is rarely effective because shape is an identity channel, not an ordered channel.
    • Normalization of data and correct channel usage (e.g., diameter vs. area for circles) is crucial to avoid misleading visuals.
  3. Visualization Critiques:
    • Pie Chart (Religion + Ethnicity):
      • Don’t blend categories inconsistently.
      • Summation/rounding errors degrade credibility.
      • Often, a multi-level or partitioned approach (sunburst) is better than a single combined slice.
    • Bivariate Map (Wisconsin):
      • Always confirm if data needs normalization (absolute vs. percentage).
      • For circles, use area for magnitude (not diameter).
      • Binning approach for choropleth must be transparent (equal intervals, quantiles, etc.).
  4. Exam Strategy:
    • The professor consistently references Mackinlay’s ranking. It’s essential for critique: “Why is length better than area for that data?” “Why is color suboptimal for quantity or time?”
    • Summaries or references to the visualization pipeline or Fun(B)ike’s model show deeper reasoning in critiques.
    • Life factor from misusing diameter instead of area is a recurring trap in real-world charts.

Final Words

  • The lecture underscores design and task alignment. Time-oriented data requires special considerations: selecting the right temporal arrangement and scale.
  • The professor’s demonstrations show the breadth of time-based techniques, from simple line charts to advanced multi-resolution or interactive approaches.
  • In critiques, always check for fundamental correctness (no mixing of categories, consistent binning, appropriate scale channels).

These notes capture the full lecture details, each step of the recap, the time-visualization content, and the critiques. Use them for reference and deeper review.

Data Visualization, Redo
This post is licensed under CC BY 4.0 by the author.