These questions boil down to proposing potential ways to expand on this work.
Methodology
Given that Wikipedia has many users editing many articles, is there any reason why using edit data to represent political preferences in a more nuanced way would not improve the work, and perhaps make the political preference signal less noisy? My thinking is this could work by representing every user as a vector of numbers of edits with length equaling the number of articles, and then using either clustering, matrix factorization, graph embeddings, or neural networks to learn better representations of these users for predicting article quality. Thinking about political affiliation on a one-dimensional spectrum from liberal to conservative is often a useful heuristic when the dataset is small, but in this case could be an unnecessarily lossy compression of the true perspective distribution of users.
Right now, this work focuses on identifying a relationship between user perspective diversity and edit quality, reinforcing evidence posed in other work that opinion diversity leads to higher quality of discourse. But I wonder how can we transition this work from basic research to translational research for actually improving opinion diversity? In other words, are there ways we can use this data to learn about the things that drive opinion diversity, whether it's the topics of the articles being discussed, the particular tendencies of users who engage in articles edited by more diverse populations, or the types of edits that are actually added.
These questions boil down to proposing potential ways to expand on this work.
Methodology
Given that Wikipedia has many users editing many articles, is there any reason why using edit data to represent political preferences in a more nuanced way would not improve the work, and perhaps make the political preference signal less noisy? My thinking is this could work by representing every user as a vector of numbers of edits with length equaling the number of articles, and then using either clustering, matrix factorization, graph embeddings, or neural networks to learn better representations of these users for predicting article quality. Thinking about political affiliation on a one-dimensional spectrum from liberal to conservative is often a useful heuristic when the dataset is small, but in this case could be an unnecessarily lossy compression of the true perspective distribution of users.
Right now, this work focuses on identifying a relationship between user perspective diversity and edit quality, reinforcing evidence posed in other work that opinion diversity leads to higher quality of discourse. But I wonder how can we transition this work from basic research to translational research for actually improving opinion diversity? In other words, are there ways we can use this data to learn about the things that drive opinion diversity, whether it's the topics of the articles being discussed, the particular tendencies of users who engage in articles edited by more diverse populations, or the types of edits that are actually added.