Sunday - Day 1
Sunday
- 10 AM Data Science
- 11:45 AM Data Science 2
- Split up at 3 PM
- Digital humanities keynote with Miriam Posner: data vs everything (what makes something not data)
- Natural language 4 vis (better captions)
- 4:45 PM Natural language for vis (short papers)
Also saw slides for sports vis + nl4viz tutorial.
Highlights
Session 1
- Progresive data analysis (DS): adjust latency based on people's expectations. Lookup pictures later
- Motif-based dynamic network analysis - send links to Ron, linked small multiples make graph analysis tractable
Session 2
- Rebecca Faust Anteater is so cool- visualize program state over time. Specify what you care about (sort of) in advance, store program state in a DB. Kate Isaacs former PI, (now at Utah)
- Communication analysis paper... need to review more. Learned about SurVis (publish browser for BibTex). Need to analyze its techniques/would people use Datasette as compliment. Can add hover timeseries card to each faceted tag (or in DD)
- Remco Chang Keynote... a conceptual framework for describing hypothesis... what does that mean for auto vis analysis? Do we need to be stricter about what a "task" is? paper
Session 3
- Captions are what people read first... more text is better paper, 2k responses!
Session 4
- Natural language UI vis helps people edo EDA better (with Eugene Wu). Leaned on Codex model to generate structured SQL (didn't train own LLM). https://ieeevis.b-cdn.net/vis_2022/pdfs/w-nlvis-1011.pdf
Raw notes
Session 1
- Author Jean: his students made life timeline for his 50 year birthday
- Applied in PSEUDO paper one- a faster way to do graph correlation using LSH (locality sensitive hashing)
- Visual data science: biascope helps compare models that might be unfair - I don't understand yet how unfairness scores are computed
Session 2
- Compared jupyter, observable, and JS in notebook. Graph analysis seems approachable. Not sure about research impact. Andromeda demo algo.
Session 3
- Accidentally found Romain's sports viz
- Data vs everything keynote... see the gay marriage database engineering article / EW lin essay on subjectivity of data capture
- Reminder of how tricky it is to capture the subjective
- Stardog: semantic data analysis
- Switched to Nl4viz after digital humanities .
Session 4
- Klaus: GPT3 can help write captions, but you need to give it a good prompt, otherwise it makes buzzfeed 1liners, and explanations can be iffy . people sure like captions though.
- Richard Brath: Qualitative / quant vis... lets add qualititaive data to quant graphics. Find a place to squeeze the text, then choose what text to insert
- Huyen: Low code for making a wordstream graph. Use "compromise" JS library to do in-browser POS tagging. (Tweeted)
- Contextualization... was a bit blurry for me, but natural language prompt can be ambiguous (remove column can be remove encoding, or remove filter). Use surrounding data to clarify intent
- natural language 4 vis
- Local downloads paper: w-nlvis-1011.pdf
- 2 contributions: Turn NL into parametrized sql, generate interface from spec (parametrized SQL). Could picture this as being a more useful form of vis recomendation (not just 1 graph, but a tool that lets you fiddle with a hypothesis)
- Diabetes prevalance vis... 10 -> 60 percent of older adults using smartphone between 2011 - 2021. Voice input helps solve that tapping around is hard!
Backlinks