In standard machine learning, we treat data as a passive observer—a snapshot of the world that stays still while we draw a decision boundary. But in high-stakes domains like credit scoring, college admissions, or insurance, the "data points" are people with agency. When a bank publishes a model for loan approval, applicants naturally "game" the system, shifting their features to cross the threshold for a positive outcome.
This phenomenon, formalized as Strategic Classification by Moritz Hardt, Christos Papadimitriou, and colleagues in Strategic Classification, transforms learning into a game. The learner sets a classifier, and the agents respond by manipulating their features to maximize their utility.
The Knowledge Gap: The Latent Cost of Gaming
The fundamental challenge in strategic classification is that the learner rarely knows how much effort it takes for an agent to change. To build a robust model, you need to know the "cost function" of the population. If it is easy for a spammer to change their IP address but hard to change their domain age, the classifier should weigh the latter more heavily.
Early theoretical models assumed these costs were known and fixed. However, as Sagi Levanon and Nir Rosenfeld argue in Strategic Classification Made Practical, this is rarely true in practice. Costs are latent, private, and highly heterogeneous. A wealthy applicant might find it "cheaper" to improve a credit score than a low-income applicant, a disparity explored by Vijay Keswani and L. Elisa Celis in Addressing Strategic Manipulation Disparities in Fair Classification.
If the learner guesses the cost function incorrectly, the resulting classifier will be fragile. The research direction has therefore shifted toward learning under unknown or personalized manipulation costs.
Learning from Revealed Preferences
How can a model designer account for costs they cannot see? The key intuition is to treat agent behavior as a revealed preference.
In Strategic Classification from Revealed Preferences, Aaron Roth, Zhiwei Steven Wu, and their co-authors propose an online learning approach. By deploying a sequence of classifiers and observing how agents respond, the learner can "triangulate" the underlying cost functions. If an agent moves their features from $x$ to $x'$ to get a better label, they are revealing that the benefit of the label exceeds the cost of the move. Over time, the learner can minimize "Stackelberg regret"—the difference between their performance and the performance of the best possible classifier they could have chosen if they had known the costs from the start.
This idea extends to collaborative science and regulation. Michael Jordan and Stephen Bates demonstrate in Incentive-Theoretic Bayesian Inference for Collaborative Science that the mere fact that a researcher invested in a costly experiment is itself a signal of their private information, allowing regulators to bound false discovery rates without seeing the raw data.
Robustness Through Dual Norms
In many real-world settings, you don't have the luxury of multiple rounds to learn from behavior. You might only get one chance to deploy a model. In One-Shot Strategic Classification Under Unknown Costs, Elan Rosenfeld and Nir Rosenfeld show that robustness in this "one-shot" setting can be achieved through a specific form of regularization.
They prove that if the agent's cost is defined by some norm (e.g., how much they can change their features in total), the learner can remain robust by applying a penalty based on the dual norm of the classifier's weights. This mathematical bridge allows designers to build models that are "strategically-aware" even when they have zero prior data on how agents will manipulate their features.
From Gaming to Improvement
While much of the field focuses on "gaming" as a nuisance, a parallel thread asks: can we use these incentives for good?
In Causal Strategic Classification: A Tale of Two Shifts, Guy Horowitz and Nir Rosenfeld distinguish between "gaming" (changing features without changing the true label) and "improvement" (changing features that actually improve the underlying state, like taking a course to improve skills).
By understanding the causal relationship between features and outcomes, Stratis Tsirtsis and Manuel Gomez Rodriguez show in Decisions, Counterfactual Explanations and Strategic Behavior that a decision-maker can provide explanations that specifically incentivize agents to take actions that are beneficial for both the individual and the system.
Summary of the Mechanism
The shift in strategic classification research can be summarized as moving from static optimization to dynamic inference:
- Observe: See the manipulated features agents provide.
- Infer: Treat the gap between original and manipulated features as a signal of latent costs (revealed preferences).
- Adapt: Update the classifier or regularization strength to account for these costs, ensuring the model remains accurate even as agents continue to adapt.
By acknowledging that costs are unknown and heterogeneous, researchers like Han Shao and Avrim Blum in Strategic Classification under Unknown Personalized Manipulation are moving the field toward a more realistic, "practical" version of incentive-aware machine learning.