New Tools Towards Computational Democracy Studies

Abstract

This dissertation is divided into two parts. In each, I propose a separate novel computational technique informed by Social Science theories, targeted at studying modern democratic issues.

In the first part, I leverage a particularity of Social Sciences to develop a Supervised Machine Learning approach to measurement, instead of the usual unsupervised methods. While uncommon in traditional Computer Science applications, there are many phenomena in Social Sciences whose accurate empirical observation is either contentious or impossible (e.g. ideology, populism, democracy, genocide, corruption, fraud), which has made them subject to diverse attempts of subjective classification by scholars - classifications of parties, leaders, regimes, governments, etc. I show how a supervised constrained ensemble learner can be used to leverage such a specificity in order to generate measures of social phenomena that: are bounded numerical scales whose values can be inherently interpreted; that mitigate biases present in the existing subjective label sources used for training; that can be generated out-of-sample; and that can be forced to retain consistency to existing Social Science theories. To illustrate the usefulness of this technique, I employ it to evaluate democratization and backsliding in hundreds of countries over history, and also across the Swiss cantons 1919-1999 and the U.S. House districts 1900-1930.

In the second part, I propose a technique that resorts to a seldom explored sub-field of AI known as Multi-Agent Systems, to translate canonical Game Theory formal models of voting into computer simulations of elections. Simulations of electoral results are regularly used to study actual elections, but they usually come from extrapolations of past results or sampling from assumed distributions. However, electoral volatility makes the past not always a good predictor of electoral results; those methods must rely on rigid parametric assumptions; and worse, they are incapable of accounting for electors’ eventual strategic behavior. My approach tries to better approximate the data generating process of election results by simulating arbitrarily many individual virtual electors at the same time, each simultaneously making their own strategic choices about whether to turnout (strategic abstention) and who to vote for (strategic wasted-vote avoidance). I illustrate the usefulness of this method in 2 ways. First, to predict the UK 2019 election seat distribution. Second, I use the simulations to generate synthetic election data and manually apply fraud to them, to then evaluate existing electoral fraud detection tools – showing their rate of false positives and false negatives.