TU Wien:Visualisierung VU (Waldner)/Made-up exam questions

This is a collection of questions that I have made up to test my (and other's) understanding of a certain topic. These are NO PREVIOUS EXAM QUESTIONS. They are in no way affiliated with the professors! They are here to help learn and to test one's knowledge. (Note: If anyone has the time, please create proper questions for Slide 15,16,17)

My ausarbeitung can be found here: Datei:TU Wien-Visualisierung VU (Waldner) - Made-up exam questions - ausgearbeitet.pdf

SLIDE 1: INTRODUCTION

Human–Data Interaction and response-time thresholds

• Why do Miller/Nielsen-style response-time thresholds (0.1s, 1s, 10s) matter for visualization design?
• Brushing & Linking vs Panning/Zooming: Pick one and explain what „real-time“ means.
• When the system has a response time of ~500ms, which interactions still „feel“ interactive? Which start to break?

Data size, scalability, and why “better hardware” isn’t the whole story

• Compare the Head Volume (30MB) with the Microscopy Volume (>1TB). What changes fundamentally in how you visualize them.
• Why can‘t we visualize Exascale Computing Solutions?
• Give two reasons why buying a faster machine might not fix interactivity for large tasks

Definitions and the visualization pipeline

• What is the pipeline diagram of a visualization?

• If you speed up only Rendering, when will the user still feel lag?
• Explain why some efficiency strategies target rendering only, while others target mapping and rendering together.

Web-based interactivity: d3 vs WebGL

• Why can WebGL render more different individual dots in a scatterplot than d3 without hitting it‘s limit?
• You must build an interactive scatterplot with 5 million points. Which architecture would you use and what interaction compromises might you consider.

Strategy selection under constraints

• Given a scenario: interactive exploration of a global map at many zoom levels, which two strategies are most natural and why?
• What is empty space skipping and how does it help volume rendering more than in for example bar charts?
• When do aggregate visualization improve accuracy of insight, not just performance?
• What is a progressive visualization?
• What is prefetching and caching?
• What is in-situ visualization? Benefits? Drawbacks?

Slide 2: Fundamentals

Core definitions and scope

• How can we define Visualization?
• Explain, what each of these terms imply in practise: computer supported, interactive, visual representation, (abstract) data, amplify cognition

Areas

• Differentiate Scientific Visualization, Information Visualization an Visual Analytics/Visual Data Science

Resource Limits

• The slides explicitly mention limitations of computers, humans an displays. Give one concrete example.
• Most visual encodings are ineffective. Give reasons as to why.

Search Space Metaphor

• In the Search Space Metaphor graph from the slides, what do “known space”, “consideration space” and “proposal space” mean?

Four nested level of visualization design

• Explain the four nested levels Domain, Abstraction, Idom, Algorithm:
• What are the problems in these nested levels? What can go wrong in each?
• Imagine this: A team builds a super fast visualization of an Aircraft carrier for a game. But nobody uses it or knows what to do with it. Which level was one wrong?
• What questions must you ask yourself for each layer?
• What kind of decisions are being made in the abstraction layer? What asks the WHAT question?
• What kind of decisions are being made in the idom layer? What asks the WHY question?
• What kind of decisions are being made in the idom layer? What asks the HOW question?
• Explain the difference between the visual encoding idom and interaction idom.

Principles / Slogans

• Explain the slogan “No Unjustified 3D”? Give an example for unjustified and justified 3D.
• Explain the slogan “Eyes over memory” and give an example.
• Explain the slogan “Get it right in black and white”. What is this guarding against. Give a failure mode.

Data-centric techniques

• What is a scalar field visualization?
• What is volume rendering?
• What is isosurface rendering?
• For a scalar field, when would you pick isosurfaces and when would you pick volume rendering? Name a tradeoff.
• What is a vector field visualization?
• Explain why vector field visualization needs different methods than a scalar field

Slide 3: Survey

Fluidity revealing Information (Un/Foldables)

• What is an „un/foldable“ visualization?
• What makes an „un/foldable“ visualization different from a plain filter or a hard drill-down?
• Explain the focus score and effect scope
• What does „unfolding scale“ mean? Why would having more states matter?

Slide 4: WebVis Tutorial

Real-Time Web-Based Visualizations

• When would you pick a SVG, canvas or WebGL for an interactive scatterplot?
• At 300% zoom, what differences should you see?
• Fill out this chart
• What makes a Network Visualization different from a regular scatterplot of what needs to be drawn and interacted with. Which technology (SCG, canvas, WebGL, ...) would you use?

Slide 5: Cuda tutorial

CPU vs GPU

• Compare CPU computing vs GPU computing
• When can data be parallelized?
• What different types of parallelization are there?
• What is Amdahl‘s Law?
• How does CUDA work (hardware perspective)?
• If a block has 128 threads, how many warps is that? And why does the unit „warp“ matter for performance and control flow?

Slide 6: WebGPU Tutorial

Probably not important

Slide 7: Algorithms

Efficient visualization algorithms and scalability factors

• What is the difference between perceptual scalability and interactive scalability?
• Why is overplotting not just a readability issue? How can we fix this graph?

Aggregate visualizations

• Contrast a point map versus an aggregate map. What changes in the visual structure and what changes in the computational workload?
• Explain the benefits of a histogram for aggregating.
• What is a Choropleth Map?

Binning

• What is Hexagon Binning?
• Why use hexagon binning instead of square binning?
• What is being binned in geographic binning versus a Choropleth Map?
• What is the problem with binning?
• In this example, what parameters can flip just by changing the bin size from 10x10 to 30x30?
• If you had to pick a bin size for an interactive visualization, what signals would tell you that your current bin size is too coarse or too fine?

Kernel Density Estimation (KDE)

• What is a Kernel Density Estimation?
• Explain Bandwidth in KDE using the “under-smoothed vs over-smoothed“ idea. What errors do each extreme create for interpretation?
• When would you prefer binned aggregation over KDE (and vice versa)?

Plots

• What is a violin plot and what are it‘s strengths/weaknesses?
• What is a density scatterplot?

Clustering

• How is a cluster Map different from binning or KDE?
• For density-based clustering, what kind of problems/failures can occur?

Interactive Scalability

• How should you implement zooming?
• How can you encode more information instead of just zooming?

Spatial data structure for fast spatial queries

• What is a KD-Tree?
• What is a BSP-Tree
• What is a Quad-Tree?
• Compare kd-tree, BSP-tree, and quad-tree: what’s the splitting rule for each, and what breaks when the data is highly dynamic (moving points)?
• Given the “flat table” vs “multi-dimensional table” idea, what operations correspond to slicing, roll-up, and projection—and what does brushing & linking trigger in that pipeline?

Progressive Visualizations

• What does progressive visualization promise during user interaction?
• Why should we randomly pick the data-points that needs to be processed first?

Data Tiles

• What are data tiles?
• Explain the tile pyramid for zooming. What gets recomputed per zoom level? How does a quad-tree show up implicitly?
• What is the difference between raster tiles and vector tiles?
• Why does can we prefetch during „idle time“?

Slide 8: VolVis

Volumetric Data Basics

• In the voxel grid (i,j,k) with one scalar L, what does it mean to „sample“ the volume along it‘s way? What are you sampling and where do these positions come from.
• The slides list examples like density, pressure, temperature, velocity, etc. Why is velocity tricky if the slides claim one scalar value L per data point?
• What are ways to render a volume?
• Compare indirect (surface) rendering with direct (volume) rendering in terms of what gets assigned color/transparency.
• Looking at the image grid, what kinds of anatomical/context information might you lose when moving from direct rendering to segmented/label-based surface views (or vice versa)?
• When would a slice based view be preferable to a 3D DVR?

Direct Volume Rendering (DVR)

• Walk us through the DVR Pipeline (Shoot a ray -> Sampling -> Classification -> Compositing) and say WHAT data is being processed.
• In Raycasting, what controls quality and speed?
• Why are people calling raycasting „Image-order (backward mapping)“ and voxel projection „Object Order (forward Mapping)“?

Transfer functions

• What is a transfer function?
• Why is choosing a transfer function an iterative process? What does „find boundaries by thresholding“ mean in practise?
• Compare MIP, average, first-hit, isosurface-like, cut, and full compositing: What does each one do in the context of transfer functions?
• In alpha compositioning, what does it mean for opacity to „accumulate along a ray“?

Real-Time DVR & Parallel rendering

• What are the primary problems of real-time DVR?
• The slides define 2 requirements for parallel processing: Separable and streamable. Use raycasting to explain both.

Spatial data structures

• Why is an octree described as a 3D generalization of a quadtree, and what does “regularly divide into eight equal octants” imply for adaptivity?
• The bricking slides list two purposes: fit in GPU and visibility culling. Explain the link. How does bricking enable culling?
• Compare octree (fixed subdivision) versus multi-resolution bricking (arbitrary subdivision). What tradeoff do you expect in efficiency?

Visibility Management

• Distinguish view-frustum culling, occlusion culling, blackface culling, and empty space culling
• Why does the slide say “opacity of a voxel is view-independent” but “opacity of a pixel is accumulated along a ray (view-dependent)”?

Empty Space Skipping

• How can we skip empty space without looking inside a node to determine if it is really empty?
• What extra power do histograms give over min-max?
• Why might a KD-Tree be better at empty space skipping than a octree?

Out of Core Rendering

• What is out of core rendering?
• Define „working set“ in bricked single-pass rendering. Also explain why it‘s view dependent and output sensitive.
• What problem does the „address translation from virtual space to brick cache“ solve?
• What happens, when the working set is too large for even the brick cache?
• Ray guided rendering determines the working set on the GPU. Why is that attractive compared to CPU driven requests? What is the catch?
• If a brick is missing, the slide suggests „substitute with a lower resolution“. Explain how octrees make this substitution easy.
• Compare LRU and MRU in regards caching in this setting. When might MRU beat LRU for interactive navigation?

Real-time Volumetric Illumination

• The slides claim that illumination helps understand structure, size, and depth. Especially for diagnosis planning, Explain WHY those tasks benefit more than, say, a casual volume screenshot.
• Contrast local ambient occlusion versus dynamic ambient occlusion in terms of computation, precomputation , and what information is sacrificed.

Slide 9

Out-of-core visualization & Multiscale & LOD

• If a volume dataset does not fit into GPU memory, what are concrete tactics to get around that?
• Why is „Not all sub-volumes need to be rendered in full resolution“ a valid design choice?

Post-hoc & In-situ

• Define post-hoc vs in-situ processing and give a practical consequence for storage and I/O.
• What does steering mean? What does it add to the pipeline?

Coupling

• Explain tight (inline), loose (in transit), and hybrid coupling.
• What is a problem of tight coupling?
• What is the problem of loose coupling
• In this image, would you use tight or loose coupling? And Why?
• When should tight and loose coupling be used?

Parallelism

• What does „streamable“ and „separable“ mean:
• Compare data parallelism vs functional/task parallelism vs temporal parallelism: what caps scalability in each one?
• „send data vs send geometry vs send images“. Where would you use each of those.

Visibility Management

• What does „sort-last“ actually sort?
• When do use Depth (Z-Buffer) compositioning vs Alpha (RGBA buffer) compositioning?
• Why does „bricking“ matter specifically for sort-last volume rendering?
• In the slides, they mention that the „Over“ operator is associative. What is associativity so important for parallel computing?
• Compare full-frame vs sparse merging. What drives communication cost in each, and what new bottleneck can spare merge introduce?
• Compare Direct-Send vs Binary-Tree compositioning. Interpret the cost n(n-1) vs log(n)
• Why do tiled displays naturally match sort-first rendering? What gets assigned to nodes?
• Why does sort-first need screen-space bounding boxes? And why do large primitives break performance?
• How can you balance the nodes so that everyone has roughly the same workload?

Slide 10: Visual Text Analytics

Immersive analysis

• Engagement vs immersion vs embodied interaction. What is each one and what would a concrete example that fits one but not the others.
• The slides list „Natural Input -> Intuitive interaction“. Where can this backfire?

New Opportunities in immersive analytics

• Explain situated analytics with your own words. Give an example where linking data to physical objects changes the analysis outcome.
• Why might „spatial arrangement of information“ help (or hurt) analytic work compared to a traditional desktop.
• The slides mention multisensory presentation. What is that? Name an analytic task that can benefit and one where it only adds nose.

Hardware and immersive interfaces

• Which hardware capabilities are „foundational“ for immersive analytics, and what breaks if one is missing?
• Compare Fish Tank VR, CAVE and HMDs in terms of viewpoint control and collaboration potential.

Applications

• Where can HMDs be used?
• Across medical, molecular, flow, climate, and sports tactics examples—what common need makes immersive environments attractive?
• Choose one application (e.g. Flow, Climate, ...) and explain what „being inside the data“ means that a 2D desktop struggles with.
• When is 3D justified?

Embodied interaction

• Explain the core idea of ImAxes and what new operations it enables compared to mouse-based axis controls.
• Tilt Map: Why would „change mapping based on viewing angle“ be useful, and what is a danger of angle-dependent encoding?
• What are Scaptics and how can they be used?
• What are input modalities for immersive analytics?

Network data

• “Networks have no inherent dimensionality.” So what is the actual argument for trying true 3D network layouts in immersive systems?
• Summarize the reported benefits and caveats of immersive 3D visualizations for abstract data and give a plausible reason for each caviat.

Platforms

• Room-sized vs Table-sized vs Egocentric. What are those and what are the benefits?
• What are trade-offs of immersive visualizations?
• What’s the conceptual shift when moving from VR analytics to AR analytics?
• Differentiate embedded visualizations vs embedded physicalization, and give an example of each.

Slide 11: Visual Text Analytics

Text information and text analysis

• Why is “text information is unstructured” not just a storage/format issue but a visualization problem?
• What is lexical text analysis?
• What is syntactic text analysis?
• What is semantic text analysis?

Single Document Visual Analysis

• In a word cloud, why does „frequency = important“ break down?
• Wordl/Word clouds can not be compared between different documents. What would you change so that two word clouds CAN be compared?

Corpus-level structure

• What is term co-occurrence?
• Explain the term-document matrix and cosine similarity in your own words.
• How does a similarity matrix turn into a 2D Map?
• In corpus maps, what can go wrong if the proximity is interpreted too literally?
• How can we specify what words are important in some document?

Word embeddings

• What are synonyms?
• What is Polysemy?
• Bag-of-words struggles with synonyms and polysemy. How do embeddings help, and what do they still lose compared to context-aware models?

Slide 12: MolVis

Is this even relevant?

Real-Time Rendering Tricks

• What is an impostor?
• Why do „standard impostors“ and simple 2D billboards break down in dense molecular senses? And what does „correct depth output“ fix?
• What are procedural impostors?

Scaling to huge models

• What artifacts can appear at LOD transitions and how would you reduce them?
• Explain the occlusion culling pipeline. Why render previously visible molecules first, and then build a hierarchical Z-buffer with mip levels?

Illumination, Color, and Visibility Management

In dense data, why are cutaway views and “which proteins are visible?” a hard problem. What makes it an optimization problem in multi-instance settings?

Slide 13: Spatial Temporal Data

Since tomorrow is exam and I covered most of it already, I will keep this section very short...

Statistical Attributes on Maps

• Why is it so important in Choropleth Maps to normalize the values to make them relative? Give a concrete example where absolute counts produce a misleading conclusion.
• „Choosing the correct mapping“ what is the difference between sequential and diverging color maps? And what property of the data tells you which one to pick?
• Spatial resolution: explain how changing the spatial aggregation unit (coarser vs finer) can flip the story a choropleth tells, even if the underlying raw data didn’t change.

Connections

• Why does drawing all connections on OD-relations as individual edges quickly break down? What can we do?
• Explain the idea of “bundle edges based on forces” in your own words and name one parameter/design choice that would strongly affect the outcome.

Slide 14: Gen AI

Automatic Chart generation

• Compare chart recommendation systems „from data“ vs “from natural language”. What inputs do they need, and what are typical failure modes for each?

Big-picture map of “LLMs in Data Visualization”

• The slides split the topic into: generating visualizations, understanding visualizations, and workflow integration. What does what mean? What are the problems? What formats are recommended?

NL2Vis

• In the NL2Vis pipeline diagram, where do you see the „translation boundary“ that most often causes errors?
• Token limits is a major problem. How would you do to process a big dataset?
• What can go wrong with chunking, context engineering, and aggregation?
• Explain why machine readable formats like SQL or XML help more than a list of columns.
• Failure analysis says that most failures are with Data rather than with Visualizations. Why? Give two examples.
• What can fail during graph creation, when an LLM creates a visualization?
• The prompt-iteration strategy self-repair outperforms role-playing. Why is that?

LLM Visualization Literacy

• Define LLM Visualization Literacy
• What is the VLAT/Mini-VLAT Test?
• What can Vision Language Models do what LLMs can‘t?
• The error types (where LLMs (and sometimes humans) are failing) are grouped into data, encoding, and reasoning. Give one example each, and explain which type you expect to increase most when charts use nested marks or layered encodings

LLMs inside workflows

• What is human in the loop?
• The slides list LLM roles: connector, simulator, programmer, assistant. Pick one role and describe a realistic scenario where that role helps without replacing the human analyst.
• Give reasons why AI collaboration might fail

Slide 15: 3D Terrain Visualization

Probably not relevant. But please add questions

Slide 16: Flood Simulation

Very probably not relevant. But please add questions

Slide 17: Large Complex Graphs

Very relevant. But I ran out of time, so here are a few AI generated questions. If someone has time, PLEASE add meaningful questions and remove the bad ones here:

Foundations

• A graph vis decision can be framed as What? Why? How? Pick a concrete analysis goal (e.g., “find clusters” vs “trace paths”) and argue how it changes (a) the target (nodes/edges/paths/groups) and (b) the encoding choice (node-link vs matrix vs hybrid).
• Given a messy node-link diagram, choose two interaction categories from the slide (Select/Explore/Reconfigure/Encode/Abstract/Filter/Connect) and explain what each would let you do that directly addresses the mess.

Graph basics & terminology + data structures

• Using the terminology slides, explain the difference between a path, a circle, a clique, and connected components—and give one quick “what might I use it for?” per term in analysis.
• You have a graph with millions of nodes and you repeatedly ask: “who are the neighbors of v?” and occasionally: “is there an edge (u,v)?” Compare adjacency matrix vs adjacency list vs edge list for these queries.

Complex graphs as sets & hypergraphs

• When would you prefer an Euler diagram over a Venn diagram for set-typed data, and what do the aesthetic principles (well-matchedness / well-formedness / area proportionality) buy you?
• The slides highlight why many-set Venn/Euler diagrams don’t scale. Explain the 2^n regions issue and give two distinct failure modes that appear as n grows.

Matrix-based set/graph encodings

• The matrix slides list when matrices are useful vs when to avoid them. Give one scenario for each side (use / avoid) and justify using the exact constraints mentioned (reordering, scale, pixel resolution, audience familiarity).
• Compare a Venn/Euler view with an UpSet/OnSet-style view for many intersections: what changes in what the viewer can do (tasks) and what is lost?

Bipartite graphs + projections + hypergraph conversions

• Explain how (a) a biadjacency matrix relates to a bipartite graph, and (b) how a hypergraph can be viewed/converted into a bipartite representation.
• What is a bipartite projection, and what kind of information can it introduce or distort compared to staying bipartite?

Very large graphs

• Define node aggregation into super-nodes/meta-nodes and name the three bases for aggregation listed (hierarchies, attributes, topology). For each, give a quick example of when it’s the right choice.
• Contrast “aggregation of edges” vs “aggregation of nodes (and edges)”: what visual problem does each target, and what new ambiguity can each introduce?

Communities, clustering, biclustering

• The slides describe communities via modularity (dense inside, sparse between) and show a hierarchical layout. Explain why a hierarchy is useful even if the underlying graph is “flat.”
• What is biclustering in the context of a biadjacency matrix, and what does “rearranging rows and columns to maximize modularity” actually accomplish visually?

Edge bundling & routing for readability at scale

• Compare edge bundling vs an edge density map: what question does each answer better, and what detail does each hide?
• Edge routing can use grids/quadtree/voronoi and represent edges as polylines/Béziers. Pick one spatial structure + one curve type and argue how that combination affects (a) bundling strength and (b) geometric faithfulness.

Data tiles / multiscale pre-rendering for real-time exploration

• Explain the core idea of data tiles for graphs (multiple zoom levels) and identify what must be precomputed/stored to make panning/zooming responsive.