Poll-Confident Voters in Iterative Voting

Anaëlle WilczynskiAAAI 2019 · DOI

Abstract

Traditional voting models often assume voters possess perfect information about the preferences of others. In reality, voters rely on a mix of potentially biased public polls and truthful but limited local social interactions. This paper introduces the Poll-Confident Iterative Voting model to study how these information sources interact. We demonstrate that while social networks act as a buffer against misinformation, a strategic polling institute can still manipulate outcomes using a "specter" strategy—reporting false leads to trigger strategic shifts. Our simulations show that these dynamics typically converge and can significantly improve Condorcet efficiency, despite the theoretical complexity of the underlying stability problem.

Introduction

In large-scale elections, voters rarely know the exact intentions of the entire electorate. Instead, they aggregate information from two primary sources: global public opinion polls and local observations from their social network. While polls provide a broad view, they can be manipulated. Conversely, social networks offer truthful insights but are limited in scope. This paper addresses how voters reconcile these conflicting signals and whether a strategic actor can exploit this information gap to steer election results.

The Poll-Confident Iterative Voting Model

The authors propose a model where voters update their beliefs about candidate scores based on an initial poll and subsequent changes observed in their social network.

Belief Updates and Confidence

Voters start with scores from a public poll. They only update these scores if a neighbor in their directed social graph changes their vote. A key parameter is the pivotal threshold ($p$): a voter considers a candidate a "potential winner" only if that candidate is within $p$ votes of the current leader, according to the voter's internal (and possibly skewed) tally.

Dynamics: Local vs. Global

The model explores two types of dynamics:

  • Local Dynamics: A single initial poll is released, followed by rounds of strategic deviations until an equilibrium is reached.
  • Global Dynamics: A sequence of polls is released periodically, each time the electorate reaches a local equilibrium, prompting further strategic adjustments.

Strategic Manipulation: The "Specter" Strategy

A central contribution of this work is the analysis of a strategic polling institute. The authors developed heuristics for "Election Enforcing," where the pollster broadcasts false results to ensure a target candidate wins.

The most effective approach is the "Specter" Strategy. Rather than falsely reporting their preferred candidate $X$ as the winner, the pollster reports a "threat" candidate $Y$ (whom target voters dislike) as the leader with a slim margin over $X$. This scares voters who prefer $X$ over $Y$ into abandoning their truthful favorites to vote strategically for $X$ to prevent $Y$ from winning.

Experimental Results

The model was evaluated using synthetic datasets of 10,000 instances with 100 voters and 10 candidates, using various social network structures (Erdős-Rényi, scale-free, and homophilic).

Convergence and Condorcet Efficiency

While determining stability is theoretically NP-hard, the dynamics almost always converge in practice. Convergence is most stable in very sparse or very dense networks, with the most volatility occurring at a density of approximately 0.25. Interestingly, iterative voting generally increases Condorcet efficiency—the frequency with which a candidate who would win all head-to-head matchups is elected—as voters gain more information through the process.

Frequency of convergence (left) and Condorcet efficiency (right) across varying network densities and confidence thresholds.

Manipulation Success

The effectiveness of strategic polling is highly dependent on network density. In sparse networks, where voters have fewer local "truth" sources, the pollster's influence is strongest. The "Election Enforcing" heuristic can more than double the win frequency of a target candidate compared to truthful polling. However, as social networks become denser, they act as a "truth-filter," making it harder for the pollster to maintain a false narrative.

Election frequency for different target candidates under various manipulation and network density scenarios.

Conclusion

This research highlights the complex interplay between public information and private social signals in strategic voting. While social networks provide a degree of resilience against misinformation, strategic actors can still exert significant influence through sophisticated polling tactics like the "specter" strategy. Future work will explore more realistic voter memory models and the impact of competition between multiple polling institutes.

Summary updated

BibTeX
@inproceedings{wilczynski2019poll,
  author = {Anaëlle Wilczynski},
  title = {{Poll-Confident Voters in Iterative Voting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33012205},
}
Figure