IS6335 Chap.6 View Manipulation and Big Data Visualization
View Manipulation and Big Data Visualization
Four moves, and an order for applying them
Week five continues the table material and adds the moves that rescue a view once it is too full. The published topics name design view manipulation by arrange, facet, reduce and reorder, then put big data visualization immediately afterwards, which is the situation that forces you to use them.
The week also names the scatterplot matrix, the heatmap and the streamgraph, and its tutorial covers ggplot2 and data.table.
Each move changes something different. Arrange changes how space is used, which was the previous chapter's subject.
Facet splits one view into several by a categorical attribute, producing small multiples; the reader gains a clean within-panel comparison and loses the direct across-panel one unless the scales are shared. Reduce removes data from the view, either by filtering, which keeps some items and drops the rest, or by aggregating, which replaces many items with a single summary mark.
Reorder changes nothing about what is shown and only changes the sequence, which is why it is the cheapest improvement available and the one most often forgotten.
Filtering and aggregating fail differently
Both are reductions and they are not interchangeable. A filter can hide the outlier the reader was looking for, and nothing in the resulting view will say so.
An aggregate keeps every item in the calculation but destroys the spread, so a reasonable-looking mean can sit on top of a distribution with two peaks. The rule that follows is to name the target first: reducing is safe when the target is a trend and dangerous when the target is an outlier.
Then reduce only what the question does not need, filtering on an attribute the reader agrees is out of scope rather than on one that merely tidies the picture.
Faceting has its own condition. Share the axis scale across panels unless you state clearly that you have not, because independently scaled panels make a small series that fills its panel look like a large one.
Reordering comes last, by the value the question is about, and never alphabetically unless the reader's task is to look up a known name.
When there are more marks than pixels
The problem the course calls big data visualization is not really about file size.
Beyond some number of marks, drawing one mark per item stops working: marks land on top of each other, the densest region looks identical to a merely busy one, and a thousand overlapping points cannot be told from ten.
Three remedies exist and they differ in honesty.
Transparency is cheapest and degrades gracefully, but it remains an accident of how many marks happen to overlap and it saturates, so past a certain density everything is solid again. Binning is better, because it converts overlap into an explicit derived attribute, the count per cell, which can then be given to a proper ordered channel; this is why the heatmap and the scatterplot matrix appear in the same week as big data.
Sampling is the third route and the one that must be stated out loud, because a sampled view answers a question about the sample while the reader will assume it answers one about the data.
Streamgraphs, and the limits of a stacked shape
The streamgraph, also named in the week's topic list, applies the same idea to parts over time: many series stacked into one flowing shape.
It is excellent for seeing which band grows and useless for reading any band's value, because only the outer boundary sits on anything resembling a baseline.
It is the extreme case of the stacked bar trade-off from the previous chapter, and it is worth choosing deliberately rather than because it looks impressive.
All of this material is most easily learned on a view that has already failed, because the failure tells you which move is needed. Diagnose before you remedy.
A chart that is hard to read because sixty names are crammed onto an axis has a different fault from one that is hard to read because its values span four orders of magnitude, and the second is not fixed by any of the four manipulations at all.
What this chapter covers
- 01
Arrange, facet, reduce and reorder as four distinct moves
- 02
Filtering against aggregating, and how each one fails
- 03
Sharing axis scales across faceted panels
- 04
Reordering as the cheapest improvement available
- 05
Overplotting, and why it is not about file size
- 06
Transparency, binning and sampling as three remedies
- 07
The streamgraph and the limits of stacked shapes
Choose between reducing and rescaling
- 2Say what a random sample preserves and what it loses.
- 2Rule on the shape task and on the outlier task.
- 2Give the alternative and what has to be reported.
Key terms
- Facet
- Splitting one view into several panels by a categorical attribute. It buys a clean within-panel comparison at the cost of the across-panel one unless the scales are shared.
- Overplotting
- The state in which marks land on top of one another so that density is no longer readable. It is driven by marks per unit of space rather than by the size of the file.
- Binning
- Dividing a range into cells and counting the items in each. It converts overlap into an explicit derived attribute that an ordered channel can carry honestly.
- Streamgraph
- Many series stacked into one flowing shape over time. It shows which band grows and cannot support reading any band's value, because only its outer boundary has a baseline.
View Manipulation and Big Data Visualization FAQ
Should I reduce first or reorder first?
Reorder first, then reduce, because reordering is free and reducing is not. Sorting by the value the question is about costs nothing, changes no data and frequently answers the question on its own, since the reader wanted to know which items are at the extremes. Reducing removes items from the view, and every removal is a decision that has to be justified and reported.
Doing the free move first also tells you how much reduction is actually needed, which is usually less than it looked before the sort. Faceting comes in only when a second attribute is genuinely part of the question rather than merely present in the file.
A default filter is set on our dashboard. Is that a problem?
It is, whenever the reader can draw a comparison the filtered view cannot support. A default filter is a reduction the reader did not choose and often cannot see, so a conclusion about a change across periods gets made from a view containing one period. It gets worse when the default is the current period, because an incomplete period has a low total for reasons that have nothing to do with the measure.
Either default to a complete period, or leave the filter visible and label the current period as partial so that the shortfall is attributed correctly.
Assessment move
Practise on a view you have already built and are dissatisfied with, rather than on a fresh one. Write down what the reader is trying to find, then name the property of the view preventing it, then pick the move.
Keeping a short log of which move fixed which fault across the six lab exercises gives you a genuinely personal checklist by the time the project starts, and the log entries turn directly into the steps of analysis the report rubric asks for.
Working through View Manipulation and Big Data Visualization in IS6335? Sia is AskSia’s AI Information Technology tutor — ask any IS6335 View Manipulation and Big Data Visualization question and get a clear, step-by-step explanation grounded in how IS6335 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
