Polis: Scaling Deliberation by Mapping High Dimensional Opinion Spaces

Deliberative and participatory approaches to democracy seek to directly include citizens in decision-making and agenda-setting processes. These methods date back to the very foun-dations of democracy in Athens, where regular citizens shared the burden of governance and deliberated every major issue. However, thinkers at the time rightly believed that these meth-ods could not function beyond the scale of the city-state, or polis. Representative democracy as an innovation improved on the scalability of collective decision making, but in doing so, sacrificed the extent to which regular citizens could participate in deliberation. Modern tech-nology, including advances in computational power, machine learning algorithms, and data visualization techniques, presents aunique opportunity to scale out deliberative processes. Here we describe Polis, an open source web application capable of collecting and synthesizing feedback from people in a scalable and distributed fashion. Polis has shown itself capable of building shared understanding, disincentivizing counterproductive behavior (trolling), and cultivating points of consensus. It has done this in the context of journalistic and academic research, and directly as part of decision-making bodies at local and national levels, directly affecting legislation. These results demonstrate that deliberative processes can be scaled up beyond the constraints of in-person gatherings and small groups.Key Words: deliberation, collective intelligence, unsupervised learning, active learning.

https://gyazo.com/1272e02943155b514268445d784e94ee

Missing values are filled in by averaging per comment.

Users with few votes (<7) are removed

PCA calculates using the power iteration method, efficient because the results of previous calculations can be reused

Filling in the missing values with averages would result in those with fewer votes being plotted near the center.

So we correct it with sqrt(C/Cp), where C is the number of comments and Cp is the number of votes cast by person p

Square root is to soften

However, this scaling was disabled by comment routing (control of the order in which comments are submitted)

Because comments with high variance appear first.

opinion group

First, clustering with K=100 for load reduction

Then clustering with K=2-5

Apply a smoothing function to prevent flapping early in the conversation

How representative each comment is of the group.

Once a group is identified, we calculate how representative each comment is of the group.

Estimating the mean of the Bernoulli distribution to find the probability that someone in the group will vote and someone outside the group will vote

Skip top half of p.9

Comments are selected at weighted random for efficient use of user time

The priority formula is

Term that increases when there is a large number of affirmative votes.

Term that decreases if there are too many paths

Terms that increase with PCA-loaded comments

The more votes are cast, the more the term decreases.

multiplied by

---