Interactive views for a public that will not read the documentation

Year 3 / Module 33 Data visualisation in practice / Unit 6 / Interactive views for a public that will not read the documentation

Interactive views for a public that will not read the documentation

HCI3109 · Unit 6 of 10

Interaction lets a reader who will not read the manual find their own question, but interaction carries a cost in attention and orientation, and the evidence on public and general audiences shows that literacy and expectations decide whether the view is used.

Interactive views for a public that will not read the documentation

The brief for a public-facing visualisation usually arrives with interaction attached: a filter, a slider, a tooltip. The reasoning is sound, because a reader who will not read the documentation can only find their own question by manipulating the view. This sixth of ten units argues that the reasoning carries an obligation the brief rarely states. Interaction moves the work of finding the question onto the reader. The evidence on general audiences says that work is done only where expectation and literacy have made it worth doing.

Four conditions decide whether an interactive view is used. They are whether the reader expects anything of it, whether they hold the literacy to operate it, whether a metered connection will load it at all, and whether the question they brought is served by staying.

Interaction is the mechanism by which a reader finds their own question

Yi, Kang, Stasko and Jacko (2007) reviewed interactive visualisation systems and found existing frameworks describing low-level operations rather than reader benefits. They proposed seven categories organised around the user's intent: select, explore, reconfigure, encode, abstract or elaborate, filter, and connect.

Elmqvist and colleagues (2011) named the quality that keeps those categories usable. Working from best-in-class examples, they proposed fluid interaction, defined through direct manipulation, embodied interaction, flow, and Norman's gulfs of execution and evaluation.

Heer and Shneiderman (2012) supplied the inventory: twelve task types in three categories. The first two are data and view specification (visualise, filter, sort, derive) and view manipulation (select, navigate, coordinate, organise). The third is analysis process and provenance (record, annotate, share, guide). That third category is the one a public view omits, and it would answer the reader's first question, where to begin. Guide means leading a reader through an analysis or a story; its absence is why documentation is written and then not read.

Interaction charges the reader, and expectation decides whether the charge is paid

Lam (2008) screened 484 papers and collected 61 interaction-related usability problems reported across 32 user studies. She sorted them into seven costs derived from Norman's stages of action: decision, system-power, multiple input mode, physical-motion, visual-cluttering, view-change and state-change. Her decision cost names the step before any control is touched. A reader who does not know what the view is for cannot begin.

A reader grants attention before evaluating any design detail. Muller and colleagues (2009) named the failure display blindness. Interviews with 91 users at 11 public displays showed audiences expected uninteresting advertising and ignored the screens. Users looked at the authors' display of information for students more often than at the one carrying shop coupons.

Literacy decides what the interaction returns

If expectation decides whether the view is met, literacy decides what the encounter yields. Boy, Rensink, Bertini and Fekete (2014) proposed item response theory as the basis for assessing visualisation literacy. They provided two tests for line graphs, later extended to bar charts and scatterplots. Lee, Kim and Kwon (2017) built the instrument out: 53 items over 12 visualisations and eight tasks. A tryout with 191 test takers gave a reliability coefficient of 0.76, and the correlation between literacy and aptitude for learning an unfamiliar visualisation was 0.64. Chevalier and colleagues (2018) take the question into elementary school. They report what they learned from how teachers instruct and how students acquire basic visualisation principles.

Neither the public nor the paid analyst takes the interaction up

The decisive evidence is null results from the field. Boy, Detienne and Fekete (2015) ran three web-based field experiments on a popular news and opinion outlet and a popular visualisation gallery. They tested whether introductory narrative techniques increased engagement with exploratory visualisations. Against their hypotheses, adding introductory stories did not increase engagement in exploration.

The two sites are worth naming. The stimulus was the CO2 Pollution Explorer. It was published first on the visualisation gallery visualizing.org, and then in French on Mediapart, a news and opinion outlet that rarely runs interactive graphics. The authors assigned new browser connections to a version with an introductory story or to one without. About four thousand unique connections arrived between January and June 2014, and a second explorer drew about thirteen hundred in a single weekend. The denominator let the authors separate readers who sought the view out from readers who met it in a news feed. The story made no measurable difference to exploration in either group.

Readers employed to work with data behave the same way. Kandel, Paepcke, Hellerstein and Heer (2012) interviewed 35 data analysts across 25 organisations in healthcare, retail, marketing and finance. They documented pain points, outstanding challenges and barriers to adoption.

No single interaction serves every reader, and the author's model of them is a design decision

Kim and Heer (2018) measured performance across task types and data distributions with varied cardinalities and entropies, over 12 encoding specifications of trivariate data. They found that effectiveness varies with both, which extends existing models of encoding effectiveness. Coloured scatterplots perform well for comparing individual points, but poorly for summary tasks as categories multiply.

Bostock, Ogievetsky and Heer (2011) bound input data to arbitrary elements of the native document object model rather than behind a toolkit-specific abstraction. They reported improved expressiveness, better integration with developer tools and, for animation and interaction, substantial gains over intermediate representations. A reader using a screen reader reaches text in the document, not pixels on a canvas, so that choice decides who can reach the marks at all.

An interactive view encodes a model of its reader, and that model can be studied. Hullman (2020) surveyed 90 authors and interviewed 13 influential designers. She found a contradiction between their acknowledgement that depicting uncertainty is valuable and the norm of omitting it, which she modelled as rhetoric.

The contested question: does interaction produce insight for a general audience, or only for a motivated one?

The first position holds that interaction is the mechanism of insight. Yi and colleagues (2007) built their taxonomy around the benefits readers obtain. Elmqvist and colleagues (2011) treat interaction as the catalyst for the dialogue with the data. Heer and Shneiderman (2012) make fluency the criterion of a good tool. Ynnerman, Lowgren and Tibell (2018) extend the claim to public communication, and they coin exploranation for the convergence of exploratory and explanatory visualisation.

The second holds that the benefit has been shown on motivated populations and abstract tasks, while the field evidence is about non-use. Muller and colleagues (2009) locate the decision before the interaction, in expectation. Lam (2008) enumerates the costs, beginning with forming a goal. Boy and colleagues (2015) added an on-ramp and exploration did not increase. Kandel and colleagues (2012) find employed analysts routing around their tools.

Both can be right, because they measure different populations and outcomes, so the disagreement is empirical. What would settle it is a study most of this literature has not run. A pre-registered field study of a public view would report its denominator, so that readers who never engaged are visible, as Muller and colleagues (2009) reported theirs. Such a study would measure three things. First, whether a reader reaches a correct answer to a question they brought rather than one the study set. Second, whether they can state the case against it, which separates use from agreement. Third, the cost in time and errors against a static alternative.

The two positions rest on different evidence. The first rests on taxonomies and best-in-class examples (Yi and colleagues, 2007; Elmqvist and colleagues, 2011; Heer and Shneiderman, 2012). The second rests on field experiments that logged roughly five thousand connections (Boy, Detienne and Fekete, 2015) and on interviews with 91 display users and 35 analysts (Muller and colleagues, 2009; Kandel and colleagues, 2012). While the dispute is open, a designer should report the denominator of a public view so non-use is visible, and add a guide that answers where to begin. The strongest objection is that every field study here measures engagement rather than insight, so the claim that a public reader finds their own question remains a prediction.

Where this is heading

Two directions are visible. The first is the convergence Ynnerman and colleagues (2018) name as exploranation, in which a view is built to be both explained and explored. The second is measurement. Visualisation literacy is now instrumented (Boy and colleagues, 2014; Lee, Kim and Kwon, 2017). Authors' assumptions about their readers can be recovered from the authors themselves (Hullman, 2020). Chevalier and colleagues (2018) point at the same conclusion, that a chart's audience is made rather than found. The open question is whether public visualisation will adopt the reporting habits of the field studies that document non-use.

The sources this unit rests on

Each note states what the cited work did and found and what it adds to the argument. Identifiers and metadata were checked on 2 November 2026.

Bostock, M., Ogievetsky, V. and Heer, J. (2011). D3 Data-Driven Documents. IEEE Transactions on Visualization and Computer Graphics, 17(12), 2301-2309. doi:10.1109/tvcg.2011.185 Presented a representation-transparent approach that binds input data to arbitrary elements of the native document object model, reporting improved expressiveness and animation and interaction with substantial performance gains. This is why interactive public views became cheap to build.
Boy, J., Rensink, R. A., Bertini, E. and Fekete, J. D. (2014). A Principled Way of Assessing Visualization Literacy. IEEE Transactions on Visualization and Computer Graphics, 20(12), 1963-1972. doi:10.1109/tvcg.2014.2346984 Proposed item response theory as the basis for assessing visualisation literacy, with two tested instruments for line graphs later extended to bar charts and scatterplots. Gives a measurable account of the skill an interactive view assumes.
Boy, J., Detienne, F. and Fekete, J. D. (2015). Storytelling in Information Visualizations. Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 1449-1458. doi:10.1145/2702123.2702452 Ran three web-based field experiments on a popular news and opinion outlet and a popular visualisation gallery. Against the authors' hypotheses, adding introductory stories did not increase engagement in exploration, the unit's strongest field evidence that an on-ramp does not produce use.
Chevalier, F., Henry Riche, N., Alper, B., Plaisant, C., Boy, J. and Fekete, J. D. (2018). Observations and Reflections on Visualization Literacy in Elementary School. IEEE Computer Graphics and Applications, 38(3), 21-29. doi:10.1109/mcg.2018.032421650 Reported lessons from how teachers instruct and students acquire basic visualisation principles in elementary school, and proposed research directions. Supports the claim that literacy is formed, not assumed.
Elmqvist, N., Vande Moere, A., Jetter, H. C., Cernea, D., Reiterer, H. and Jankun-Kelly, T. J. (2011). Fluid interaction for information visualization. Information Visualization, 10(4), 327-340. doi:10.1177/1473871611413180 Collected best-in-class interactive visualisations and proposed fluid interaction as an operational definition grounded in direct manipulation, embodied interaction, flow, and the gulfs of execution and evaluation. Argues that interaction is what turns a display into a dialogue.
Heer, J. and Shneiderman, B. (2012). Interactive dynamics for visual analysis. Communications of the ACM, 55(4), 45-54. doi:10.1145/2133806.2133821 Set out a taxonomy of twelve task types in three categories, the third being analysis process and provenance: record, annotate, share and guide. Guide is the part a public view typically omits.
Hullman, J. (2020). Why Authors Don't Visualise Uncertainty (the published title is American). IEEE Transactions on Visualization and Computer Graphics, 26(1), 130-139. doi:10.1109/tvcg.2019.2934287 Surveyed 90 authors who create visualisations for others and interviewed 13 influential designers, finding a contradiction between their stated valuation of depicting uncertainty and the norm of omitting it, and proposed a rhetorical model of the omission. Shows that an author's model of the reader is recoverable.
Kandel, S., Paepcke, A., Hellerstein, J. M. and Heer, J. (2012). Enterprise Data Analysis and Visualization: An Interview Study. IEEE Transactions on Visualization and Computer Graphics, 18(12), 2917-2926. doi:10.1109/tvcg.2012.219 Interviewed 35 data analysts from 25 organisations in healthcare, retail, marketing and finance, documenting pain points, outstanding challenges and barriers to adoption. Establishes that even paid readers work around analysis tools.
Kim, Y. and Heer, J. (2018). Assessing Effects of Task and Data Distribution on the Effectiveness of Visual Encodings. Computer Graphics Forum, 37(3), 157-167. doi:10.1111/cgf.13409 Measured performance across tasks and data distributions over 12 encoding specifications of trivariate data, finding effectiveness varies with both, with coloured scatterplots good for individual comparisons but poor for summary tasks as categories grow. Shows no single design suits an unknown reader's task.
Lam, H. (2008). A Framework of Interaction Costs in Information Visualization. IEEE Transactions on Visualization and Computer Graphics, 14(6), 1149-1156. doi:10.1109/tvcg.2008.109 Collected 61 interaction-related usability problems from 32 user studies out of 484 screened papers and placed them into seven costs derived from Norman's stages of action, with decision cost a gulf of goal formation. Supplies the unit's account of what interaction charges the reader.
Lee, S., Kim, S. H. and Kwon, B. C. (2017). VLAT: Development of a Visualization Literacy Assessment Test. IEEE Transactions on Visualization and Computer Graphics, 23(1), 551-560. doi:10.1109/tvcg.2016.2598920 Developed a literacy assessment test with 53 items over 12 visualisations and eight tasks, tried out on 191 takers, with a reliability coefficient of 0.76 and a 0.64 correlation between literacy and aptitude for learning an unfamiliar visualisation. Makes the reader's entry requirement measurable.
Muller, J., Wilmsmann, D., Exeler, J., Buzeck, M., Schmidt, A., Jay, T. and Kruger, A. (2009). Display Blindness: The Effect of Expectations on Attention towards Digital Signage. Pervasive Computing (Pervasive 2009), Lecture Notes in Computer Science 5538, 1-8. doi:10.1007/978-3-642-01516-8_1 Ran two studies: interviews with 91 users at 11 public displays showed audiences expected advertising and ignored the screens, with the authors' student display looked at more often than their coupon display, and repertory grid interviews with 17 users. Shows that expectation, not design quality, gates attention.
Yi, J. S., Kang, Y. A., Stasko, J. and Jacko, J. A. (2007). Toward a Deeper Understanding of the Role of Interaction in Information Visualization. IEEE Transactions on Visualization and Computer Graphics, 13(6), 1224-1231. doi:10.1109/tvcg.2007.70515 Reviewed interactive visualisation systems, found existing frameworks described low-level operations rather than benefits, and proposed seven intent-based categories: select, explore, reconfigure, encode, abstract or elaborate, filter, connect. Reframes interaction as reader intentions.
Ynnerman, A., Lowgren, J. and Tibell, L. (2018). Exploranation: A New Science Communication Paradigm. IEEE Computer Graphics and Applications, 38(3), 13-20. doi:10.1109/mcg.2018.032421649 Coined exploranation for the convergence of exploratory and explanatory visualisation in communication, in which exploration methods communicate results and explanation methods enrich exploration. Names the direction public views are moving.

Check your understanding & apply it · 4 questions, 1 essay

1A public interactive view receives about five thousand visits, but almost nobody interacts with it. Which diagnosis does the unit's evidence support?

2A designer is choosing controls for a public view that a reader will meet for under a minute. Using the unit's sources, which judgement is defensible?

3A team claims interaction lets every reader find their own question, so an interactive public view will produce insight for a general audience. Using the unit's evidence, which prediction is best supported?

4A colleague argues: 'Interaction always helps, because a reader who will not read the documentation can only find their own question by manipulating the view.' Which judgement does the unit's reading of the evidence support?

5Essay A city transport authority wants a web page where residents can explore last year's bus punctuality. The data cover 1,200 routes with a median of 40 scheduled journeys each, and the press office wants readers to conclude that performance is broadly good and that the worst routes are isolated problems. The page will be read on phones by people who arrive from a news story and stay for under a minute. Using this unit's sources, say what interactive controls you would provide and what you would refuse to provide, how you would serve the readers who never interact at all, predict before release what readers are most likely to take from the page wrongly, name the one check you would run, and then give the strongest counter-argument to your approach and answer it.

I would start from the reader's intent rather than from the controls, because Yi, Kang, Stasko and Jacko (2007) built their seven categories around what a reader is trying to do, and Lam (2008) shows that decision cost, the step of forming a goal, is charged before any control is touched. So the page opens with one default view: 1,200 routes as a ranked distribution of on-time percentage, with the 40-journey median stated, which is a question a reader can answer without acting. I would provide three controls only, each mapping to one intent: a filter by region, a search by route number, and a toggle between route-level values and the distribution. I would refuse a free-form expression field, linked multi-view brushing and any animated transition, because Heer and Shneiderman (2012) place coordination and derivation far along the analytic dialogue, and a reader with under a minute will not pay their cost. I would also refuse, on the press office's instruction, to show only the best-performing regions: the framing the office wants is a claim about selection, and omitting the rest would enter the page through that omission in the way Hullman (2020) documents for uncertainty. For readers who never interact, the page must answer the press office's question in static form before the controls appear, because Boy, Detienne and Fekete (2015) found in three field experiments that adding an introductory story to an exploratory view did not increase exploration, so I cannot assume an on-ramp will convert a non-interacting reader. Muller and colleagues (2009) reported the denominator of their display study, 91 users at 11 displays, and I would do the same: log how many readers see the controls at all, and treat that number as a finding. Before release I would predict two errors, both drawn from the sources rather than from intuition. The first is that readers take the map of worst routes as evidence about the network, when a route with 40 journeys produces unstable rates by construction. The second is that readers without the literacy to operate the toggle never see the distribution and read the ranking as the whole picture, which is the entry requirement that Boy, Rensink, Bertini and Fekete (2014) and Lee, Kim and Kwon (2017) make measurable rather than assumed. The one check I would run is a short task-based session with readers recruited from the news traffic rather than from the authority's staff, in which each participant is asked for the question they arrived with, then for their answer, then for the strongest argument against it, with the time taken recorded; Kim and Heer (2018) show that encoding effectiveness depends on the task and the data distribution, so the check must test my readers' tasks and not my own. The strongest counter-argument is that interaction is exactly what a 1,200-route dataset needs, because no single static figure can carry that many cases, and that refusing controls privileges the press office's summary over the resident who wants their own street. That argument is right about the data and is the reason the default view must show the whole distribution rather than an average. It is wrong only if it assumes the reader will interact; the check above measures whether they do, and if they do not, the static distribution is doing the work and the controls are decoration.