Draft the pipeline architecture for Polis-like tools #53
Description
Now that we've essentially got an MVP that can reproduce Polis math outputs verbatim (minus base-clusters and associated kmeans downsampling code), I've been thinking about how to describe the generalize pipeline of a Polis-like tool, the unit of experimentation. This feels helpful for conceptualizing the opportunity for new interfaces like my Opinion Landscape Painter and related exploration/gamification dynamics.
- data collection
- voting: reaction of participants as attraction/neutrality/repulsion toward content
- submitting new statements: new content to which other participants can react
- filling in missing data
- Polis: column-wise per-statement mean imputation
- Alternatives: TabPFN frontier tabular data model imputation, perhaps trained on other human preference datasets.
- dimensional reduction.
- project participants in lower-dimensional 2D
- Polis: reducer = PCA. fit and transform participants via linear relationships.
- Opinion Landscape Painter: PCA, PaCMAP, LocalMAP, more?
- project statements in lower-dimensional 2D [optional]
- Polis: reducer = PCA. transform statements with model fit to participants.
- Opinion Landscape Painter: n/a (only PCA allows this without rethink)
- project participants in lower-dimensional 2D
- participant labeling (aka grouping)
- Polis: KMeans clustering.
- Opinion Landscape Painter: manually assigned groups. auto-assigned group based on other clustering algos later?
- vote-able statement selection (statement presented for voting)
- Polis: comment routing. relies on labels.
- Opinion Landscape Painter: n/a
- statement statistics on groups
- Polis: group representative statements. differences between groups.
- Polis: group-aware consensus statements. sameness across groups.
- Opinion Landscape Painter: dynamic group representative statements.
- group storytelling/sensemaking/identification.
- Polis: ???
- Polis v2: LLMs group summaries?
- Opinion Landscape Painter: descriptive group labels with share links.
To Dos
- create flowchart of high-level abstractions
- create flowchart of low-level programmatic abstraction
- list ideas for diversity/experimentation at each step