Catch Me if I Can: Detecting Strategic Behaviour in Peer Assessment

Ivan Stelmakh, Nihar B. Shah, Aarti Singh

Abstract

Peer assessment is a scalable grading method but vulnerable to strategic manipulation, where participants rank others unfairly to boost their own standing. We present a principled, non-parametric statistical test to detect such behavior at an aggregate level. Using a novel dataset of human strategic patterns collected from a gamified experiment, we demonstrate that our test effectively controls false alarms and maintains high detection power, particularly when supplemented with a small amount of supervised data. This work provides a tool for system designers to decide when to switch to more robust, albeit less efficient, impartial mechanisms.

Introduction

In competitive environments such as MOOCs, university courses graded on a curve, or academic peer reviews, participants often face a conflict of interest. If a reviewer is graded against their peers, they have a mathematical incentive to assign lower marks to others to improve their own relative standing. While "impartial" aggregation rules exist to mitigate this, they are often less data-efficient and can discourage reviewer effort. To determine if the trade-off for impartiality is necessary, organizers need reliable methods to detect if strategic behavior is occurring. Previous research has focused on rating-based systems; however, detecting manipulation in ranking-based systems is significantly more complex, as the optimal cheating strategy depends on the specific subset of peers assigned to a reviewer.

Detection Methodology

The authors propose a non-parametric statistical test, designated as Test 1, based on the intuition that a strategic reviewer's ranking will likely result in a better outcome for their own work than a random ranking would.

  1. The Test Statistic ($\tau$): For each reviewer, the test calculates the "impact" of their ranking on their own submission's final position. It compares the actual position of their work under the submitted ranking to the expected position if they had submitted a ranking sampled uniformly at random.
  2. Permutation-based Null Distribution: To ensure statistical significance, the authors construct a null distribution by permuting authorship and conflict matrices. This simulates a scenario where rankings are independent of the reviewers' own submissions.
  3. Supervision: The test can operate in an "unsupervised" mode (using only peer rankings) or a "supervised" mode (incorporating a small set of impartial rankings from instructors or non-conflicted third parties). Supervision significantly enhances the test's power.

Experimental Setup and the CMU Dataset

To elicit real-world strategic behavior, the authors conducted a gamified experiment with 55 graduate students at Carnegie Mellon University. In a 5-round game, students acted as "authors" of submissions with hidden values and were required to rank four peers. While the first four rounds involved "truthful bots" to teach mechanics, the final round involved direct competition between students. Analysis of students' reasoning revealed six clusters of human strategies, such as "Distance" (ranking others based on their proximity in quality to the reviewer) and "Reverse" (inverting the true quality order). This dataset serves as a benchmark for human behavior under strategic incentives.

Strategic Behavior in Rankings

Unlike rating systems where a simple "zero-out" strategy suffices, ranking systems require a more nuanced approach. The authors demonstrate that there is no single dominant strategy. The most effective approach is often the "Distance" strategy, where a reviewer places their closest competitors at the bottom of their list.

Expected gain from various ranking strategies based on the reviewer's ground-truth position.

As shown in the figure above, the expected gain from different strategies varies significantly depending on the reviewer's own ground-truth position. For instance, a reviewer in the second position can gain significantly from certain strategies, whereas a reviewer in the fourth position may actually see a negative expected gain from the same strategies.

Key Results

The proposed test demonstrates several critical properties:

  • False Alarm Control: The authors provide mathematical proof that the test reliably controls the false alarm rate at a pre-specified level (e.g., $\alpha = 0.05$).
  • Detection Power: Empirical evaluations using the CMU dataset showed that the supervised version of the test could achieve a 100% detection rate in certain rounds. Even without supervision, the test maintained non-trivial detection power.
  • Robustness and Scalability: Simulations indicate the test remains effective despite "noisy" reviewers or violations of the assumption that better authors are better graders. Furthermore, the test is computationally efficient, capable of processing thousands of reviewers in minutes.

Conclusion

While the current test is highly effective at detecting successful aggregate manipulation, it has limitations. It is not designed to identify specific individuals, which avoids the ethical complications of individual accusations but limits targeted intervention. Furthermore, it may not flag unsuccessful attempts where strategies cancel each other out. Future work will focus on detecting unsuccessful manipulation attempts and integrating these detection tools into real-time peer-grading platforms to provide instructors with actionable data.

Summary updated

BibTeX
@inproceedings{stelmakh2021catch,
  author = {Ivan Stelmakh and Nihar B. Shah and Aarti Singh},
  title = {{Catch Me if I Can: Detecting Strategic Behaviour in Peer Assessment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i6.16611},
}
Figure