[Data Visualization] Reduce and Focus+Context

Summary: Interacting with Views

• Changing a single view: encoding, arrangement, order
  • Viewpoint: pan, zoom (geometric/semantic), rotate
  • Data: slice, cut, and project
• Multiple coordinated (or faceted) views
  • Juxtaposition (multiform, overview/detail, small multiples)
  • Partition
  • Superimpose (static/dynamic)

The Big Picture

• Datasets are often large and complex so drawing all of them into a single static view can be overwhelming.
• What options are available to handle such complexity? There are five.
  • 1. Change a view over time (previous lecture)
  • 1. Derive new data (Lecture 3)
  • 1. Facet data by partitioning into multiple juxtaposed views (previous lecture)
  • 1. Reduce the amount of data (this lecture)
  • 1. Embed focus and context information within a single view (this lecture)

Why Reduce?

• Reducing the amount of data shown in a view is an obvious way to reduce its visual complexity.
• Filtering simply eliminates elements (i.e., an item or an attribute).
  • Easy to understand
  • Out of sight, out of mind
• Aggregation creates a new element that stands for multiple others that it replaces.
  • Safer but cannot convey all omitted information

Filter

• Filtering reduces the number of elements shown: some elements are simply eliminated.
• Dynamic queries: tight coupling between visual encoding and interaction
  • The user can immediately see the results of the intervention
  • A display for showing a visual encoding of the dataset + filter controls
• Filter items
• Filter attributes

Dynamic Queries

• Widgets:
  • Sliders
  • Buttons
  • Comboboxes
  • Text fields

Example: FlimFinder

Scented Widgets

• Standard widgets for filtering controls can be augmented by concisely visually encoding information about the dataset.
• Scented widgets
• High information density

Example: DOSFA

• Dimensional Ordering, Spacing, and Filtering Approach
• 215 attributes representing word counts
• Dimensions are ordered by similarity and filtered by similarity and importance thresholds

Aggregation

• In aggregation, a group of elements is represented by a new derived element that stands in for the entire group.
  • Many-to-one visual paradigm
• With derived attributes: average, min, max, count, and sum
• Users should be able to change the level of aggregation interactively.

Example: Histograms

• How many bins?
• What is the range of each bin? (nice numbers)

Example: Continuous Scatterplots

Examples: Boxplots and Vase Plots

• Visualization for an aggregate statistical summary of values
• Median (50% point), lower and upper quartiles (25% and 75%), and upper and lower fences (upper + 1.5 IQR and lower – 1.5 IQR where IQR = val(75%) – val(25%))
• (n) normal, (s) skewed, (k) peaked, (mm) multimodal

Examples: Hierarchical Parallel Coordinates

• Parallel coordinates for clusters
• 230,000 items and 8 attributes

Example: Multiclass Density Maps

Example: Dimensionality Reduction

• Dimensionality reduction is the transformation of data from a highdimensional space into a low-dimensional space so that the lowdimensional representation retains some meaningful properties of the original data.
• e.g., if two points are closer to each other in the original high-dimensional space, let’s project them to similar locations in 2D.

• Principal Component Analysis (PCA)
• Multidimensional Scaling
• Autoencoder
• LargeVis
• t-Stochastic Neighbor Embedding (t-SNE)
• Uniform Manifold Approximation and Projection (UMAP)
• $\dots$

• t-SNE on the MNIST dataset

t-Stochastic Neighbor Embedding

Embed: Focus+Context

• Focus: detailed information about a selected set
• Context: overview information about more of the data
• Three design choices: elide (생략하다), superimpose, distort

Why Embed?

• The limitation of geometric zooming: when zoomed in, only a small part of world space is visible.
• Focus+Context idioms attempt to support orientation by providing contextual information intended to act as recognizable landmarks.
• Overview+Detail: spatial separation
• Zoom: temporal separation
• Focus+Context: seamless focus in context
  • From “A Review of Overview+Detail, Zooming, and Focus+Context Interfaces”

• Overview+Detail

• Focus+Context

Elide

• Elision: some items are omitted from the view completely, in a form of dynamic filtering. Other items are summarized using dynamic aggregation, and only the focus items are shown in detail.
  • Different from filtering. If you filter out items or attributes, they are completely gone.

Example: DOITrees Revisited

• Multiple foci to show an elided version of a 600,000 node tree.
• Triangles: aggregate representation showing the size of sub tress.
• The focus nodes can be chosen by clicking and searching

Superimpose

• Superimposition: the focus layer is limited to a local region on a global layer that stretches across the entire view.
  • i.e., the lens metaphor

Example: TopicLens

• Global layer: DR result of documents (vis papers)
• Local layer: fine-grained DR result of focused documents

Distort

• Many focus+context idioms integrate focus and context into a single view using geometric distortion of the contextual regions to make room for the details in the focus regions.
• Is there only a single region of focus, or does the idiom allow multiple foci?
• Is the shape of the focus a radial, rectangular, or a completely arbitrary shape?
• Is the extent of the focus global across the entire image, or constrained to just a local region?

Example: Fisheye Lens

• The fisheye lens distortion idiom uses a single focus with local extent and radial shape and the interaction metaphor of a draggable lens on top of the main view.

Costs and Benefits: Distortion

• Benefits: combine focus and context information in a single view
• Costs:
• Length judgements are severely impaired (this is why geometric distortion is widely-used for node-link diagrams)
• Not familiar (the user can misunderstand the underlying structure)
• Less object constancy (need to mentally track changing objects)

Design Variants

Summary: Reduce and Focus+Context

• Reduce: reduce the number of items or attributes to be visualized.
  • Filter: simply eliminate items.
  • Aggregation: draw an element that represents a group of items.
• Embed: show the focused items but with contextual information
  • Elide: the context information is summarized.
  • Superimpose: the context information is shown as a background.
  • Distortion: the context information is geometrically distorted.
• Focus+Context vs Filtering vs Overview+Detail

Summary: Interaction

• In short, we cannot show all the details about the data a single view.
• Therefore, visualization systems should be interactive.

1. Allow the user to change the view interactively (Manipulate)
2. Derive new data
3. Partitioning into multiple juxtaposed views (Facet)
4. Reduce the amount of data shown (Reduce)
5. Show the focused data with context information (Embed)

What is the Trend?

• Most recent InfoVis system papers incorporate at least one derivation technique to summarize the entire data.
  • Clustering, dimensionality reduction, attribution score, data quality, $\dots$
• Manipulate is a must.
• Reduce (filter and aggregation) is a must.
• Facet is frequently used.
• In contrast, Embed is used less frequently.
  • Overview+Detail seems more popular than Focus+Context