For full text search please use the '?' prefix. e.g. ? Onboarding

Thursday - Day 4

Thursday (24 papers!)

  1. VIS Full Papers: ML for VIS
  2. VIS Full Papers: Digital Humanities, e-Commerce, and Engineering - Enrico Bertini
  3. VIS Full Papers: Storytelling
  4. VIS Full Papers: Provenance and Guidance - Alvitta Ottley

Highlights

  • Morning P1
    • 3 taxonomies post survey framework for AI <> Viz intersections (evaluation, generation, NL, copiloting)
      • 3rd paper taxonomy probably most tied to immediate future work, could be paper group idea (find top impacting papers to prioritize?)
    • Sketch of how to do DL dashboard design training (feature space + reward function based on insight function): Dashbot
  • Morning P2
    • Case studies: 2500 yrs of history, what-if analysis / SHAP plots for how online sales promos work
    • Traveler: how would we use linked views on parallel traces? Link 2-layer Gantt with Topomap linked view
    • contrails: TIL alpha-shape (generalized form of Sagar's Convex hull analysis (Edelsbrunner 1992))
    • Privacy-preserving vis (motivates looking back into differential privacy + rebrushing on basics of bayesian networks)
  • Afternoon P1
    • BDFW: "tutorial" / gamelike experience for engaging someone in a vis
    • Erato: DL can help with interactive storytelling / generating supporting captions + viz
    • Roslingifier, Geostorylines main demos
  • Afternoon P4: Guidance/Recommendations
    • Visualize provenance: it can help someone not get lost during EDA. History or coverage.
    • Provectories, Tradeoff (wheat) vis: summmarize log in embedding or with frequency vis
    • RUM guidance: if someone is frustrated, can you nudge a "visual fix" / next step transform without getting to clippy stage?
    • Medley: make recommendations based on INTENT (implicitly by other graphs, explicit by requesting certain compare types) - similar to storytelling sesion intent requests
    • GeVitRec: DataRecon prioritize color and positional alignment using domain specific info, not just solo one off graphs. Compare to Showme, Draco, Voyager
      • An observability specific recommender may outperform what a generic Draco alg does if we pick good heuristics

...what we demonstrate here is how a design space can be used to inject domain expertise into recommendation systems. We hope that our work makes headway on this “chicken and egg” problem, motivating further work on automating the construction of a VPDS by showing the practical benefits that can be obtained from having one available.

Raw Notes

ML for Vis

3 survey papers on intersection between #dataviz and #machine-learning (Private)

  • AI4VIS: Survey on Artificial Intelligence Approaches for Data Visualization - Aoyu Wu
  • DL4SciVis: A State-of-the-Art Survey on Deep Learning for Scientific Visualization
    • https://arxiv.org/pdf/2204.06504.pdf
    • Chaoli Wang (HK) - no paper browser after
    • Surveyed multiple journals, Conferences, excludes all non-DL based ML
    • Surveyed 59 papers! Partitioned by data type, + included tasks. Came up with a 6 dimensional taxonomy
    • Prediction
      • People go after CV because it's easy and popular... next wave is switch from datagen to vis gen!
  • A Survey on ML4VIS: Applying Machine Learning Advances to Data Visualization (Qianwen Wang et al) #to-read
    • https://arxiv.org/pdf/2012.00467.pdf (19 pages!)
    • Covers 97 papers: https://ml4vis.github.io/
    • Figure out gaps in VIS needs, strengths ML offers, and try to pair!
    • Notice combos that are frequent , and what areas are avoided or poor pairings (e.g. semi supervised leraning, or regression in wrong tasks)
    • QW on paper 1 vs 2
      • them: concept of visualization as an important data type, including how these data exist and is studied by the researchers.
      • us: shaping the needs/problems in visualization as ML task so that research can identify suitable visualization problems and employ proper ML techniques
    • Stages: Data, vis (creating, insight, style), user (interaction user profiling, vis decoding). Good use of annotation!

Next ones focused on applications -> this region was a bit trickier to follow.

  • Reinforcement Learning for Load-balanced Parallel Particle Tracing
    • https://ieeevis.b-cdn.net/vis_2022/paper_images/v-tvcg-9706326.png
    • Use reinforcement learning to efficiently split work for a supercomputer problem (16k processors). paritcle tracing used for 3d/atmosphere vis
    • Agents communicate- share/staeal/split work, estimate costs to try and split work dynamically rather than needing to anticipate good worksplits in advance
  • IDLat: An Importance-Driven Latent Generation Method for Scientific Data
    • https://ieeevis.b-cdn.net/vis_2022/pdfs/v-full-1018.pdf
    • Idea: when projecting high dim space down to latent space - can't weight data by "important" regions, regenerating latent space is hard
    • Generate latent space in a way that takes "regions of interest" into account (unsupervised approaches). They contribute a way to use "Spatial importance maps" to take that input into account. used to help with understanding Hurricane Isabel dataset.
  • DashBot: Insight-Driven Dashboard Generation Based on Deep Reinforcement Learning #to-read . Zhejiang uni sponsored
    • thumb
    • https://ieeevis.b-cdn.net/vis_2022/pdfs/v-full-1033.pdf
    • "agents" with reward fucntions for creating dashboards using domain-speciifc heuristics / reinforcement learning. Doesn't depend on large labeled corpus of datasets. use Markov Decision process.
    • Use Tableau + powerBI training data to parametrize dashboard design
    • Recommend diversity, insights, avoid diminishing returns
    • Recommended followup charts. (Focuse on penguins / cars / IMDB datasets)
    • People like the idea of rcommendation to "having key columns"
    • Analytics group: https://zjuvag.org/ . Previous github: https://github.com/zjuidg

Digital Humanities, e-Commerce, and Engineering

Digital Humanities

  • CohortVA: A Visual Analytic System for Interactive Exploration of Cohorts based on Historical Data #to-read

Commerce

Engineering (Physical, digital, Privacy

Storytelling

Provenance and Guidance

To browse further later

Missed

- Carlolabe: document browser: https://virtual.ieeevis.org/year/2022/paper_v-cga-9238399.html
- Legalvis
- Text boundary visualization
- Temporal merge tree for big heatmaps
    - https://virtual.ieeevis.org/year/2022/paper_v-full-1051.html
- Prompt-maker IDE (Sent to Amelia Wattenberger)
Backlinks