Automatic Generation of Semantic Icon Encodings for Visualizations by Vidya Setlur
- examples of using icons for data points on visualizations
- why use icons? Represent semantics, easier to parse spatially and visually, better aesthetics, but takes time to find or create icons; can we automate it?
- context establishment->query construction and image retrieval->global feature clustering; discover icons with semantic meaning to the data then cluster by similarity of visual representation
- used Mechanical Turk to evaluate automated icon set versus hand-picked, the automated set were preferred for general media concepts but less for domain specific sets or in a narrow area where auto generated icons were too realistic and hard to tell apart; Turkers like round icons
- Q: surprised that word net algorithm made such good choices for semantics; how? 'symbol' is very well represented in WordNet and Google has good search semantics to instrument
- Q: in other domains would it be this good? send the data; it works without manual intervention unless domain too narrow or hard to represent
- Q: plan to look at user-guided vs. automated or user-generated? good idea
Task-Driven Evaluation of Aggregation in Time Series Visualization by Danielle Albers
- certain visualizations are better for specific visualization tasks, but not strong guidelines available
- visual aggregation tasks: summary info in display regions; point vs summary comparisons; avoid visual clutter
- design variables: visual, mapping, computational
- compared tasks using Mechanical Turk; monthly sales data set with controlled variability; did find correlations; eg position better at point, color at summary, and box plots better than line charts for minima (surprised the researchers); users can find insights in complex data sets using visualizations
- Q: can we overlay things and get one chart for all tasks? likely not, but insights in the better choices to make in design; multiple views supports the many micro tasks
Dive In! Enabling Progressive Loading for Real-Time Navigation of Data Visualizations by Michael Glueck
- interaction transaction: user provides input and system responds; stepped vs continuous interactions
- current data exploration uses mostly stepped interactions because of large sets
- Splash Framework: explore, continuous, and instantaneous; progressively load data; well known but little used; progressive loading is hard to do right; most research scientists are not expert programmers; abstract levels of detail; improves performance in use; works with existing visualizations
- Splash Aggregator->Data Transport->Splash Cache
- compared progressive vs non-progressive representations on coarse, global, and fine features; progressive is roughly equivalent in all cases and better in most low bandwidth scenarios (except fine features); helps maintain “real-time” interaction (<200ms)
- detail can be tuned by data curator, and researchers were successful at choosing a good tuning level
- Q: what are the limitations? simplified level of detail settings, no semantic hierarchies, no dynamic elements, however supports many data types
Sample-Oriented Task-Driven Visualizations: Allowing Users to Make Better, More Confident Decisions by Niven Ferreira
- uncertainty is often ignored in visualizations; adds complexity to tasks and experts have trouble interpreting it; how can we improve understanding of uncertainty?
- important tasks: retrieve data, find extrema, find range; incorporating uncertainty: how likely? sorting and ranking tasks as well
- design guidelines: easy to interpret, consistency across tasks, stability across sample sizes, minimize visual noise; simple, light-weight, annotations, smooth transitions, natural representations
- eg use color to overlay the uncertainties when selecting a bar; use pie chart to represent likelihood of extrema; probability of ranking in list
- tested how these encodings worked in real-world analysis; 7 participants, 75 questions, measure answer + timing + confidence; all were comfortable, didn't have to think much, rank/sort was complex, sample size greatly influenced confidence, not significantly slower when using them, we're not more accurate however with these simple questions (easy to guess), correctness and confidence correlated