TL;DR
This paper introduces a new family of measures to quantify "divisiveness" in collective decision-making, identifying issues that cause disagreement among individuals. They also propose methods to control and manipulate this divisiveness, offering insights into how to foster more unified opinions or strategically influence collective outcomes.
What Problem Does This Paper Address?
In many collective decision-making scenarios, like voting or choosing among policy options, we aim to aggregate individual preferences into a single, collectively preferred outcome. Traditional rank aggregation methods focus on finding consensus – identifying the most agreed-upon options. However, this paper tackles the flip side: how to identify issues that divide a population. Understanding divisiveness is crucial for analyzing social and political phenomena, as divisive issues can lead to polarization and hinder effective collective action. Existing measures of polarization often focus on the overall distribution of opinions rather than pinpointing specific issues that cause disagreement. This work bridges that gap by developing quantifiable measures of issue-specific divisiveness.
Methods
The core contribution is a new family of divisiveness measures for rank aggregation. These measures quantify how much an issue divides a population by comparing its collective score (based on a chosen aggregation function like Borda or Copeland) across different sub-populations.
- Divisiveness Definition: The divisiveness of an issue
afor a sub-populationXis the absolute difference in its score betweenXand its complementN\X. - Alpha-Divisiveness: They generalize this by introducing an
α-parameter that allows for weighted averages across all possible sub-populations, effectively controlling the influence of sub-population size. This parameter allows their measure to interpolate between existing polarization measures (like Esteban and Ray's) and detecting minority disagreements. - Rank-Variance: As a related concept, they also define and analyze "rank-variance," which measures the spread of an issue's rank across individual preferences.
- Control Mechanisms: The paper explores two ways to "control" divisiveness:
- Removing Pairwise Comparisons: Investigating how incomplete rankings affect divisiveness measures.
- Adding Fake Rankings (INJECT): Proposing a heuristic algorithm that adds controlled "fake" rankings to manipulate an issue's divisiveness score, aiming to make a target issue either more or less divisive.
Datasets
Not discussed in the paper. The experiments were conducted using synthetic preference profiles generated by standard models like the Impartial Culture (IC) and the Urn model (UM10, UM50) with varying degrees of correlation.
Key Insights and Ideas
- Divisiveness is Distinct from Polarization: The paper demonstrates that their divisiveness measures are not always correlated with traditional polarization measures like rank-variance, especially as the number of issues increases. This highlights the value of focusing on issue-specific disagreements.
- Controllable Divisiveness: The research shows that divisiveness is not a fixed property but can be influenced. This has implications for understanding how information spreads and how collective opinions can be shaped.
- Sub-population Identification: The proposed measures can identify the specific sub-population that is most divided on a given issue, offering granular insights into societal disagreements.
- Robustness to Incomplete Information: While removing pairwise comparisons can significantly alter divisiveness rankings, especially with many issues, it suggests that eliciting a substantial portion of preferences is crucial for accurate measurement.
- Strategic Manipulation is Possible: The INJECT algorithm demonstrates that adding even a moderate percentage of controlled rankings can significantly alter the divisiveness landscape, highlighting potential vulnerabilities in crowdsourced or online decision-making platforms.
Key Results
- Divisiveness Measures: The paper provides a formal definition of a family of divisiveness measures and analyzes their properties, including bounds and behavior in limit cases (e.g., unanimous vs. fully polarized profiles).
- Computational Complexity: They show that finding the sub-population that maximizes divisiveness for the Borda score can be done in polynomial time, but this remains an open problem for the Copeland score.
- Experimental Findings: Simulations using synthetic data indicate that:
- Divisiveness measures correlate with rank-variance for a small number of issues, but this correlation decreases significantly as the number of issues grows.
- Removing as little as 10-20% of pairwise comparisons can substantially decrease the accuracy of divisiveness rankings when there are many issues.
- The INJECT algorithm can make the least divisive issue the most divisive by adding approximately 35% new agents in some scenarios. More correlated profiles (UM50) are harder to control.
Limitations and Future Work
- Limitations:
- The INJECT algorithm for control is a heuristic, and its efficiency (number of agents needed) can vary.
- The study of control mechanisms primarily focuses on a=0 divisiveness, and the computational complexity of finding maximally divided sub-populations for the Copeland score is an open problem.
- The experiments rely on synthetic data, and real-world validation would be beneficial.
- Future Work:
- Developing divisiveness measures that are more robust to external control (e.g., manipulation by fake agents or incomplete data).
- Extending the use of divisiveness measures to compare populations rather than just issues within a population.
- Exploring axiomatic characterizations of divisiveness, a common approach in social choice theory.
- Investigating the computational complexity of control mechanisms for other scoring rules.
Summary updated