In fields like medicine or autonomous driving, making decisions based on "average" outcomes can be dangerous. If a treatment helps most people slightly but causes a severe reaction in a few, a policy that only minimizes expected cost might still recommend it.
This challenge is compounded when we only have observational data—records of what happened in the past. In these datasets, the "past policy" (why a doctor chose treatment A over B) is often unknown and biased. For example, a specific drug might have only been given to older patients in the past. This creates a "lack of overlap" in the data: we simply don't know what would have happened if a younger patient had received that drug. Standard methods often struggle or produce high-variance estimates in these "data deserts."
This paper introduces a method to learn decision-making policies from historical data that prioritize avoiding high-cost "worst-case" outcomes rather than just optimizing for the average. By leveraging the conformal prediction framework, the authors provide a statistically valid "safety certificate" for every decision, even when the historical data is biased or lacks examples for certain groups.
A Robust Policy for the Tail End
The authors propose a Robust Policy ($\pi_\alpha$) that targets the "tail" of the cost distribution. Instead of minimizing the mean cost, they aim to minimize the $(1-\alpha)$ quantile (e.g., the 80th percentile of cost). Key technical components include:
- Conformal Prediction: The core innovation is using conformal prediction to construct a limit, $y_\alpha(z)$, for each decision. This limit acts as a guarantee: with a probability of at least $1-\alpha$, the actual cost will not exceed this limit.
- Distribution-Free Validity: Unlike standard regression, this method provides statistical guarantees that hold for finite samples without assuming a specific shape for the noise or cost distribution.
- Unsupervised Weighting: When the past policy is unknown, the authors use generative models (like Gaussian Mixtures or Autoencoders) to estimate the "propensity" of past decisions. This allows them to re-weight the data to simulate what would happen under a new policy.
Handling Data Deserts
Most off-policy methods fail when there is no overlap (i.e., when a treatment was never given to a certain group). This method handles it gracefully: it doesn't try to guess. If the model encounters a situation where it has no relevant historical data, the conformal limit automatically converges to the maximum possible cost.
This forces the robust policy to avoid making "blind" decisions in areas of the feature space it hasn't explored, naturally leading the policy to prefer options it actually has evidence for.
Key Results and Insights
The researchers evaluated their method on synthetic patient data (simulating blood pressure treatments) and the Infant Health and Development Program (IHDP) dataset, a well-known real-world benchmark in causal inference.
- Safety Certificates: Every decision comes with a "certificate" ($y_\alpha(z)$). This is highly interpretable for practitioners (e.g., "We recommend Treatment A, and we are 80% sure the patient's blood pressure won't rise by more than 5 points"). In both synthetic and real-world tests, the empirical probability of the cost exceeding the predicted limit was consistently close to the target $\alpha$, proving the certificate is reliable.
- Superior Tail Performance: Compared to standard linear policies that minimize expected cost, the robust policy significantly reduced the prevalence of high-cost outcomes.
In scenarios where the cost distribution is skewed or has "fat tails," the robust policy effectively trades a slightly worse average outcome for a significantly lower risk of a catastrophe.
- Computational Efficiency: The algorithm relies on sorting and interval halving, resulting in a runtime complexity of $O(n \log n)$, making it practical for large datasets.
Limitations
While powerful, the method assumes that the range of possible costs is known or has a defined maximum. If the potential "worst case" is infinite or unknown, the uninformative limit cannot be set. Additionally, while the method is "distribution-free" regarding the costs, it still relies on the accuracy of the unsupervised generative models used to estimate feature distributions. Finally, if the policy is deployed on a population significantly different from the training data, the safety guarantees may no longer hold.
Summary updated