Research map / Learning theory and algorithms / Prediction and decision losses
Statistical Collusion by Collectives on Learning Platforms
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Develops a finite-sample statistical framework letting collectives compute computable lower bounds on their success (signal planting, unplanting, erasing) on learning platforms, validated on a synthetic car-evaluation dataset.
Problem
How collectives that alter their submitted data can a priori estimate and guarantee their influence on a learning platform using only quantities inferable from their own pooled data, under finite samples and unknown parameters.
Approach
Extends algorithmic collective action to a statistical setting: the collective pools its data, uses Hoeffding concentration to derive lower bounds on its success that are computable from observed data, and adds a new objective (signal unplanting) with an adaptive label-estimation strategy.
Key results
- Lower bounds on success are below observed test success; for signal unplanting the bound suggests about 10% of agents are needed while about 3% sufficed in practice.
- Adaptive unplanting strategy with n_e=2,000 outperforms naive fixed-label planting strategies in the lower bound.
- Infinite-data limit bounds are tighter than prior algorithmic collective action bounds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase the collective size in absolute terms while holding its fraction of the population fixed | Feature-label signal planting · Vehicle evaluation, signal planting | success lower bound | Helped |
| use an adaptive label selected from pooled data rather than a fixed naive label | Signal unplanting · Vehicle evaluation, signal unplanting | success lower bound | Helped |
| set the estimation subset size to n_e = 2,000 | Adaptive signal unplanting · Vehicle evaluation, signal unplanting | success lower bound | Helped |
| use feature-only rather than feature-label signal planting (ablation) | Signal planting · Vehicle evaluation, signal planting | test-time success | Hurt |
| use the adaptive strategy with n_e = 2,000 rather than naive fixed-label strategies (ablation) | Signal unplanting · Vehicle evaluation, signal unplanting | success lower bound | Helped |
| increase n_e from small values toward a balanced estimation subset (ablation) | Adaptive signal unplanting · Vehicle evaluation, signal unplanting | success lower bound | Helped |
| make n_e overly large (ablation) | Adaptive signal unplanting · Vehicle evaluation, signal unplanting | success lower bound | Hurt |
| use the feature-label strategy to plant a signal targeting Excellent | Feature-label signal planting · Vehicle evaluation, signal planting (target label Excellent) | test-time success | No clear effect |
| compare the adaptive strategy with the naive strategy of planting Good (ablation) | Signal unplanting · Vehicle evaluation, signal unplanting | success lower bound | Mixed |
| use the proposed infinite-data bounds instead of the earlier population-level bounds | Signal planting · Infinite data regime | lower bound | Helped |
Surprising findings
- Collective influence depends on absolute collective size, not only relative size, so larger platforms may be more vulnerable.
- Empirical success was reached with about 3-5% of agents while the lower bound suggested about 10%.
Limitations
- Hoeffding inequality can be loose; bounds are conservative.
- Assumes finite feature space and identically distributed consumers.
- Upper bounds from reversed concentration are often trivial.
Experimental setup
Datasets: synthetic car evaluation dataset (3,000,000 instances)
Benchmarks: synthetic car evaluation dataset
Metrics: lower bound on collective success test-time success rate signal planting and unplanting success
Baselines: algorithmic collective action bounds (infinite data regime) naive signal unplanting strategies
Builds on: algorithmic collective action strategic classification Hoeffding's inequality
Closest papers by meaning
- A Statistical Framework for Algorithmic Collective Action with Multiple Collectives (NeurIPS 2026 · similarity 0.87)
- The Role of Learning Algorithms in Collective Action (ICML 2024 · similarity 0.86)
- Look-Ahead Reasoning on Learning Platforms (NeurIPS 2025 · similarity 0.83)
- DICE: Data Influence Cascade in Decentralized Learning (ICLR 2025 · similarity 0.83)
- Adversarial Cooperative Rationalization: The Risk of Spurious Correlations in Even Clean Datasets (ICML 2025 · similarity 0.83)
- Automating Data Annotation under Strategic Human Agents: Risks and Potential Solutions (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.