Test-time Collective Prediction

Celestine Mendler-Dünner, Wenshuo Guo, Stephen Bates, Michael Jordan

In the ideal world of machine learning, we would pool all available data into one massive "super-model." In reality, data is often siloed across different organizations due to privacy concerns, legal restrictions, or competitive interests. While Federated Learning addresses this by coordinating training across machines, it requires significant synchronization and agreement on model architectures.

This paper tackles a more flexible scenario: Collective Prediction. Imagine several parties (e.g., different hospitals or banks) each have their own independently trained models. They want to collaborate to make better predictions on a new test point, but they aren't willing to hand over their training data or their proprietary model weights. The challenge becomes how to combine these "black-box" models effectively, especially when some might be experts in one area of the data space but completely unreliable in another.

The DeGroot Consensus Model

To solve this, the authors look to social science—specifically the DeGroot consensus model, a mathematical framework for how a group of humans reaches a consensus through repeated opinion sharing.

The proposed algorithm, DeGroot Aggregation, operates in three main steps:

  1. Local Validation (The "Trust" Phase): When a new test point $x'$ arrives, each agent looks at their own small, local "neighborhood" of data points similar to $x'$. They then "query" the other agents' models to see how well they perform on this local neighborhood.
  2. Mutual Trust Scores: Based on this local performance, each agent $i$ assigns a trust score to agent $j$. If agent $j$ predicts agent $i$'s local data well, they earn more trust.
  3. Iterative Consensus: The agents then engage in a digital "discourse." They share their initial predictions and repeatedly update them by taking a weighted average of their peers' opinions using the trust scores. Mathematically, this is equivalent to finding the stationary distribution of a Markov chain.

The paper also introduces a decentralized Jackknife procedure, which allows the group to estimate the "error bars" or sensitivity of their collective prediction—essentially asking, "How much would our answer change if we kicked one expert out of the room?"

Key Insights: Adaptivity is Everything

Unlike standard "bagging" or "stacking," which often use fixed weights for models, DeGroot Aggregation is point-adaptive. The weights change depending on where the test point sits in the input space. A model that is "trusted" for predicting house prices in suburbs might be ignored when predicting prices in the city.

Agent weights adapt dynamically to regions where their individual Mean Squared Error (MSE) is lowest.

The authors prove that in the "large-sample limit," their decentralized approach actually recovers inverse-MSE weighting. This is the mathematically optimal way to combine independent, unbiased estimators. Furthermore, most ensemble methods require a separate "hold-out" validation set to figure out which models are good. DeGroot Aggregation uses the agents' existing local training data to perform this validation on the fly.

Empirical Results

The authors evaluated their method using a mix of synthetic and real-world datasets, including Boston Housing, Abalone, and financial risk prediction benchmarks.

Performance of DeGroot Aggregation showing (a) massive MSE reduction compared to simple averaging, (b) rapid convergence during consensus finding, and (c) uncertainty estimation via the Jackknife procedure.

In synthetic experiments, DeGroot Aggregation achieved a 50x reduction in Mean Squared Error (MSE) compared to simple model averaging, even outperforming an "oracle" that simply picked the single best model for each point. Across real-world datasets, DeGroot consistently outperformed standard model averaging (M-avg) and static weighting schemes. Most impressively, DeGroot often performed as well as—or better than—schemes that had access to additional shared validation data, despite DeGroot having no such extra information.

Limitations and Future Work

While highly effective, the approach has a few limitations:

  • Privacy Leaks: While the method doesn't share raw data, the "trust phase" involves querying models on local data points. This could potentially leak information about the training set, suggesting that Differential Privacy could be a necessary future layer of protection.
  • Query Costs: The system assumes agents can query each other's models an unlimited number of times. In commercial settings where API calls are monetized, the "local validation" step could become expensive.
  • Scope: The current paper focuses heavily on regression. While classification is touched upon, a more robust framework for discrete labels is a clear next step.
  • Security: While the Jackknife helps diagnose sensitivity, the consensus could potentially be "hijacked" by a group of colluding, malicious agents, an area ripe for future exploration.

Summary updated

Abstract

An increasingly common setting in machine learning involves multiple parties, each with their own data, who want to jointly make predictions on future test points. Agents wish to benefit from the collective expertise of the full set of agents to make better predictions than they would individually, but may not be willing to release labeled data or model parameters. In this work, we explore a decentralized mechanism to make collective predictions at test time, that is inspired by the literature in social science on human consensus-making. Building on a query model to facilitate information exchange among agents, our approach leverages each agent’s pre-trained model without relying on external validation, model retraining, or data pooling. A theoretical analysis shows that our approach recovers inverse mean-squared-error (MSE) weighting in the large-sample limit which is known to be the optimal way to combine independent, unbiased estimators. Empirically, we demonstrate that our scheme effectively combines models with differing quality across the input space: the proposed consensus prediction achieves significant gains over classical model averaging, and even outperforms weighted averaging schemes that have access to additional validation data. Finally, we propose a decentralized Jackknife procedure as a tool to evaluate the sensitivity of the collective predictions with respect to a single agent's opinion.
BibTeX
@inproceedings{mendlerdunner2021test,
  author = {Celestine Mendler-Dünner and Wenshuo Guo and Stephen Bates and Michael Jordan},
  title = {{Test-time Collective Prediction}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2021},
  url = {https://openreview.net/forum?id=hrkY-fe8nJV},
}
Figure