Causal Feature Discovery through Strategic Modification

Yahav Bechavod, Katrina Ligett, Zhiwei Steven Wu, Juba Ziani

Abstract

We consider an online regression setting in which individuals adapt to the regression model: arriving individuals are aware of the current model, and invest strategically in modifying their own features so as to improve the predicted score that the current model assigns to them. Such feature manipulation has been observed in various scenarios — from credit assessment to school admissions — posing a challenge for the learner. Surprisingly, we find that such strategic manipulations may in fact help the learner recover the meaningful variables — that is, the features that, when changed, affect the true label (as opposed to non-meaningful features that have no effect). We show that even simple behavior on the learner's part allows her to simultaneously i) accurately recover the meaningful features, and ii) incentivize agents to invest in these meaningful features, providing incentives for improvement.

Introduction

In high-stakes machine learning applications such as credit scoring, insurance, and college admissions, the subjects of the model are not passive data points. Instead, they are strategic agents who adapt their behavior to achieve better outcomes. This phenomenon, often termed "gaming the system," is traditionally viewed as a threat to model accuracy and stability. When agents manipulate features to improve their scores, they often break the statistical correlations that the model relies upon, leading to a "cat-and-mouse" game between the learner and the population.

This paper reframes this dynamic, suggesting that strategic manipulation is not merely a hurdle but a powerful mechanism for causal discovery. By observing how agents respond to different models, a learner can distinguish between "meaningful" features (those that causally influence the outcome) and "non-meaningful" features (proxies that are merely correlated with the outcome).

The Problem: Proxies vs. Causes

Standard machine learning assumes that the data distribution is independent of the model. In strategic environments, this assumption fails. A significant challenge arises when a model assigns weight to a proxy feature—for example, the brand of a car a student’s parents drive might correlate with academic success, but it does not cause it. If a model rewards this proxy, agents will strategically modify it (e.g., by renting a luxury car), which provides no value to the learner and eventually degrades the model's predictive power.

In contrast, meaningful features (e.g., time spent studying) actually drive the true label. The goal of the learner is to identify the true latent regression vector $\beta^*$ that represents these causal relationships, even when initial data is confounded by proxies.

Methodology: Strategic Interaction as a Natural Experiment

The authors model this interaction as an online regression problem with a feedback loop:

  1. Model Publication: The learner posts a linear regression model $\hat{\beta}$.
  2. Strategic Adaptation: An agent arrives, observes the model, and spends a budget to modify one feature to maximize their predicted score.
  3. Observation: The learner observes the modified features and the resulting true label $y$. Crucially, $y$ only changes if the agent modified a meaningful feature.
  4. Iterative Learning: The learner updates their model using Batch Least-Squares Regression on the accumulated data.

The Power of Interaction

A static learner can never distinguish a proxy from a cause if they are perfectly correlated in the training data. However, the interaction itself acts as a natural experiment. When an agent "games" a non-meaningful feature, they change the feature value while the outcome $y$ remains stagnant. This drop in correlation informs the learner that the feature is a proxy. Conversely, if an agent invests in a meaningful feature and the outcome improves, the causal link is reinforced.

Tie-Breaking for Exploration

To ensure the learner doesn't get stuck in a sub-optimal model, the paper introduces a Tie-Breaking Scheme (Algorithm 2). Because initial data may be rank-deficient, multiple models might explain the data equally well. The learner is instructed to pick a solution that puts a large weight on "unexplored" directions. This forces agents to manipulate those specific features, effectively "stress testing" them to see if they actually affect the true label.

Key Results

The theoretical analysis yields two primary guarantees:

  • Recovery Guarantees (Theorem 4.1): The learner can accurately recover the true weights ($\beta^*$) for any feature that agents have invested in sufficiently. Accuracy improves as the number of observations increases.
  • Full Convergence (Theorem 5.2): By utilizing the proposed tie-breaking scheme, the learner is guaranteed to eventually explore all $d$ features and recover the full latent regression vector $\beta^*$ accurately, provided the time horizon $T$ is sufficiently large ($T \geq dn$).

These results demonstrate that even simple, greedy behavior—re-training on past data while handling tie-breaking correctly—is sufficient for causal discovery.

Incentivizing Improvement

As the learner identifies the true causal features, the system naturally shifts from rewarding "gaming" to rewarding "improvement." Eventually, the only way for an agent to improve their score is to invest in features that actually improve their true label. This aligns the incentives of the agents with the goals of the learner, creating a system that encourages genuine progress rather than superficial manipulation.

Limitations and Future Directions

While this framework provides a robust foundation for causal discovery in strategic settings, several avenues for future work remain:

  • Linearity: The current model assumes linear relationships and linear costs; real-world interactions are often more complex.
  • Feature Attribution: The model assumes agents modify one feature at a time. If agents shift multiple features simultaneously, attributing changes in the label becomes significantly more difficult.
  • Causal Graphs: Integrating these dynamics with formal causal graph discovery, where features may affect one another, represents a promising next step for the field.

Summary updated

Abstract

We consider an online regression setting in which individuals adapt to the regression model: arriving individuals are aware of the current model, and invest strategically in modifying their own features so as to improve the predicted score that the current model assigns to them. Such feature manipulation has been observed in various scenarios -- from credit assessment to school admissions -- posing a challenge for the learner. Surprisingly, we find that such strategic manipulations may in fact help the learner recover the meaningful variables -- that is, the features that, when changed, affect the true label (as opposed to non-meaningful features that have no effect). We show that even simple behavior on the learner's part allows her to simultaneously i) accurately recover the meaningful features, and ii) incentivize agents to invest in these meaningful features, providing incentives for improvement.
BibTeX
@inproceedings{bechavod2021causal,
  author = {Yahav Bechavod and Katrina Ligett and Zhiwei Steven Wu and Juba Ziani},
  title = {{Causal Feature Discovery through Strategic Modification}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2021},
  url = {https://openreview.net/forum?id=KP6tzhspqJx},
}
Figure