Seeing Patterns, Checking Claims: Visual Analysis of Media Collections
Which view helps answer which question? Information visualization for media collections: visual encoding, linked views, uncertainty and evaluation.
A media library contains a thousand videos. A list can display every one of them, but it does not yet reveal which topics are missing or how the collection has developed. A colorful dashboard only answers these questions if its presentation fits the task.
Information visualization maps data to visible properties. Visual analytics combines such representations with computational methods and human interpretation. For our example library, the aim is to understand a collection and verify individual claims about it.
The question comes before the chart
An editorial team may want to identify underrepresented topics. Students look for related lectures. A research analysis investigates changes over several years. These tasks can use the same source data and still need different views.
Data can be encoded through position, length, area, color and other visual properties. Which encoding is appropriate depends on the data type and task. Categories need a different representation from quantitative values. Perspective and occlusion can make unnecessary three-dimensional representations harder to read. Munzner, 2014: Visualization Analysis and Design
For our library, we propose the following mapping:
| Question | Possible representation | What else must be visible |
|---|---|---|
| How many videos are there per topic? | Sorted bars | Counting rules and multiple assignments |
| How does a topic develop? | Time series | Time period and reference measure |
| Which topics occur together? | Matrix | Meaning and strength of the relationship |
| Which videos underlie the result? | Table with links | Sources and filter state |
These are design proposals. A table may be better than a chart for precise individual values. A network view would make sense only if the relationship structure itself is part of the task.
An overview needs paths to the details
Shneiderman’s well-known interaction principle begins with an overview, then supports zooming and filtering, and provides details on demand. His original paper connects different data types with typical exploration tasks. Shneiderman, 1996: The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations
In our example, selecting a topic could restrict both a timeline and the video list. A notable bar would then lead to the records used to calculate it. The selection remains visible and can be cleared.
This traceability matters. Without it, the editorial team might interpret a low value as a gap in the collection when a language filter is simply active. A view should therefore indicate which part of the collection it currently represents and which content is excluded.
Absolute values can tell the wrong story
An illustrative example: in the first year, a topic has ten videos out of 100 in total. In the second year, it has 15 out of 300. The absolute number rises, while the share falls from ten to five percent.
Which representation makes sense depends on the question. The count matters for the amount of available material. The share can help assess the distribution of editorial attention. Making both values accessible prevents a single metric from suggesting an inappropriate interpretation.
The counting rule must also be established. A video with three topics may appear three times in topic frequencies. Their sum will then no longer equal the number of videos. This relationship belongs alongside the view, not solely in technical documentation.
Automatic groups are a hypothesis to investigate
In our design, a model could group similar videos. The editorial team then checks whether the groups make sense within the subject area. Such a group is initially a result of the selected method; it should not appear as an authoritative subject category without review.
The same applies to spatial representations of learned embeddings. When many dimensions are reduced to two, the visible arrangement depends on the method and its settings. In our example, we would therefore retain direct access to videos and their features rather than inferring subject relationships from distances in the image alone.
Model evaluation in Part 6 and statement provenance in Part 4 provide important information for this. Review status can be shown as text or a symbol and must not be conveyed by color alone.
Evaluate at the right level
Munzner’s nested model distinguishes the domain problem, its abstraction into data and tasks, the visual representation and interaction, and the algorithms that implement them. Errors at an early level affect later decisions. A fast, readable graphic can therefore solve the wrong task. Munzner, 2009: A Nested Model for Visualization Design and Validation
For the library, we would accordingly test concrete questions: can the editorial team identify a deliberately introduced topic gap? Can it find the underlying videos? Does it understand the difference between count and share? Do these interactions remain responsive with a larger collection?
UX evaluation in Part 3 describes observing such tasks. In the final development project, visualizations are added only when they support an actual task involving the collection.