Conformal Candidate Certification for Offline Model-Based Optimization

Technical Summary: Conformal Candidate Certification for Offline Model-Based Optimization

Problem Statement
Offline Model-Based Optimization (MBO) seeks to identify high-performing designs $x$ (e.g., in protein engineering or materials science) using only a fixed historical dataset $D$, without further oracle access. The standard pipeline trains a surrogate model $\hat{f}_\theta$ on $D$ and optimizes it to propose candidates. However, this process inherently drives candidates into out-of-distribution (OOD) regions where the surrogate is unreliable, leading to systematic overestimation.

Existing methods attempt to mitigate this by modifying the proposal mechanism (e.g., conservative surrogates, invariant representations, or conditional generative models). A critical gap remains: these methods do not provide a per-candidate statistical certificate that a proposed design $x^*$ satisfies a target threshold $\tau$ (i.e., $f(x^*) \ge \tau$) with a guaranteed confidence level. Conflating the proposal (generation) and certification (validation) tasks results in either a conservatism penalty or a lack of rigorous coverage guarantees for individual candidates.

Methodology: Conformal Candidate Certification (CCC)
The authors propose Conformal Candidate Certification (CCC), a post-hoc wrapper that attaches a calibrated one-sided lower bound to any candidate generated by an offline MBO algorithm. CCC separates the proposal task from the certification task, treating the latter as a covariate-shift conformal prediction problem.

The core components of the methodology are:

1. Covariate Shift Modeling via Gibbs Tilt:
   The paper models the proposal distribution $Q_X$ (from which candidates are drawn) as a Gibbs-tilted version of the historical data distribution $P_X$. This arises naturally from entropy-regularized surrogate maximization:
   $$q_T(x) \propto p_X(x) \exp\{\hat{f}_\theta(x)/T\}$$
   Crucially, the same surrogate $\hat{f}_\theta$ used to generate the shift also supplies the importance weights $w(x) = dQ_X/dP_X(x)$, eliminating the need for a separate density-ratio estimation step.

2. Temperature Estimation:
   The temperature parameter $T$ is estimated via moment-matching. To avoid degenerate weights when candidates are OOD relative to the validation set, the authors pool the validation set $D_{val}$ with the candidate surrogate scores to anchor the upper tail of the reference distribution.

3. Weighted Conformal Prediction (IW-CP):
   CCC employs importance-weighted conformal prediction to restore marginal validity under the covariate shift.

   • Score Function: A one-sided nonconformity score is defined as $s(x, y) = (\mu(x) - y)/\sigma(x)$, where $\mu$ and $\sigma$ are the surrogate's mean and uncertainty. This focuses calibration probability on the lower tail (overestimation).
   • Weighted Quantile: For a candidate $x^*$, the method computes a weighted quantile of the calibration scores using Gibbs-derived weights. If the weight of the test point is too high relative to the calibration set, the bound becomes vacuous (correctly indicating insufficient support).
   • Certification: The lower bound is computed as $y(x^*) = \mu(x^*) - \hat{q}^w_{1-\alpha}(x^*)\sigma(x^*)$. A candidate is certified only if $y(x^*) \ge \tau$.

Key Contributions

1. Formulation: The paper formulates post-hoc candidate certification for offline MBO as a covariate-shift conformal prediction problem, explicitly decoupling certification from proposal.
2. Gibbs-Tilted Proposal: It demonstrates that entropy-regularized surrogate maximization induces a Gibbs-tilted proposal distribution, allowing the surrogate itself to provide the necessary importance weights for weighted conformal prediction without external density estimation.
3. Theoretical Validity: The authors prove finite-sample marginal lower-bound validity ($P\{Y^* \ge y(X^*)\} \ge 1-\alpha$) for a fresh candidate from the proposal distribution, assuming oracle weights.
4. Empirical Demonstration: A synthetic study validates that CCC successfully certifies a non-trivial subset of aggressive proposals with near-zero false certifications, whereas standard (unweighted) conformal prediction collapses under covariate shift.

Experimental Results
In a controlled synthetic study involving an aggressive Boltzmann proposal pool:

• Naive Surrogate: Accepted all candidates but achieved only 41% true precision (many false positives).
• Unweighted Conformal Prediction: Ignoring the covariate shift resulted in an empirical coverage of 0.416 against a nominal 0.90 target, with a 22.7% false-acceptance rate.
• CCC: Certified 16.7% of the candidate pool. Crucially, it maintained an empirical coverage of 0.990 at the nominal 0.90 level, with a false-acceptance rate below 0.5%. The method remained conservative but non-vacuous across various nominal levels ($\alpha \in \{0.05, \dots, 0.20\}$).

Significance and Claims
The paper claims that CCC provides a principled solution to the question offline MBO has left unanswered: "Which proposed candidates are trustworthy enough to test?"

• Separation of Concerns: By wrapping any existing MBO algorithm, CCC allows the proposal mechanism to remain aggressive while ensuring that only candidates with sufficient empirical support (via a calibrated lower bound) are advanced to expensive experimental evaluation.
• Robustness to Shift: The method explicitly addresses the failure mode of standard conformal prediction under covariate shift, which is the dominant characteristic of offline MBO.
• Modest Limitations: The authors acknowledge that the finite-sample guarantee strictly holds only for oracle weights. In practice, using estimated Gibbs weights means the guarantee is not exact, though empirical results suggest the method remains conservative (maintaining coverage above the nominal level) rather than anti-conservative. The paper notes that for noisy measurements, additional assumptions are required to certify the latent objective $f(x^*)$ directly, as the current guarantee applies to the measured response $Y^*$.

In summary, CCC offers a post-hoc certification layer that transforms aggressive, potentially unreliable MBO proposals into a set of candidates with rigorous statistical guarantees, bridging the gap between offline optimization and reliable online decision-making.