Abstract
Recently, Szufa et al. [AAMAS 2020] presented a "map of elections" that visualizes a set of 800 elections generated from various statistical cultures. While similar elections are grouped together on this map, there is no obvious interpretation of the elections' positions. We provide such an interpretation by introducing four canonical "extreme" elections, acting as a compass on the map. We use them to analyze both a dataset provided by Szufa et al. and a number of real-life elections. In effect, we find a new variant of the Mallows model and show that it captures real-life scenarios particularly well.
Introduction
In social choice theory, researchers often use "maps" to visualize how different elections relate to one another. These maps typically utilize a "positionwise distance" metric: if two elections have similar candidate-ranking distributions, they are plotted close together. However, these visualizations have historically suffered from two major issues. First, they lacked orientation; while one could see that Election A was near Election B, the absolute position on the map was meaningless without "North, South, East, or West" markers. Second, there was a significant "real-world gap," as most maps were populated with synthetic data from mathematical models, leaving it unclear where actual human voting data—such as city council results or sports rankings—fit into this abstract space.
The Compass: Four Extremes of Voting
To address the lack of orientation, we treat elections as frequency matrices (specifically bistochastic matrices), where each entry represents the fraction of voters who ranked a specific candidate in a specific position. We define four "extreme" matrices to act as the corners of our map:
- ID (Identity): Represents perfect agreement; every voter ranks candidates in the exact same order.
- UN (Uniformity): Represents a total lack of agreement; every candidate is ranked in every position equally often (Impartial Culture).
- ST (Stratification): Voters agree on the "big picture" (e.g., a specific set of candidates is better than another) but disagree entirely on the rankings within those groups.
- AN (Antagonism): Represents extreme polarization; half the voters rank candidates in one order, and the other half ranks them in the exact opposite order.
We use the Earth Mover’s Distance (EMD) to calculate the distance between these matrices. Crucially, we provide an algorithm to recover a representative election from any point on the map, justifying the use of matrices as a proxy for complex voting data.
Mapping Real-World Data
Our study utilizes a massive variety of data, including 800 synthetic elections (Urn, Mallows, Single-Crossing) and real-world preference data from PrefLib (e.g., Glasgow and Aspen city council elections). We also introduced new datasets derived from sports results, such as the Tour de France and figure skating.
By overlaying our compass, we find that real-life elections are far from random. They cluster tightly near the path between ID and UN, rarely appearing in the "Antagonistic" or "Stratified" regions. Political elections and surveys tend to sit in the middle of the ID-UN axis, while sports elections shift closer to the ID (perfect agreement) pole.
The Mallows Model and rel-phi
Among synthetic models, the Mallows model—which assumes voters have a "central" preference but make some "errors"—is the most accurate at mimicking real-life data. However, the standard Mallows model uses a dispersion parameter ($\phi$) that behaves inconsistently as the number of candidates changes.
We propose a new normalized parameter, rel-$\phi$ (relative expected number of swaps), which allows researchers to generate comparable elections regardless of the number of candidates. We identified specific rel-phi values that best simulate real-world scenarios: approximately 0.375 for political elections and 0.267 for sports.
Conclusion and Future Work
The "compass" provides a necessary frame of reference for the map of elections, revealing that human preferences occupy a specific, non-random quadrant of the voting universe. While our analysis focused on elections with 10 candidates, future work will involve proving the "Diameter Conjecture" (that ID and UN are the most distant possible points) for any number of candidates and extending this approach to non-ordinal voting systems.
Summary updated