Learn then Test: Calibrating Predictive Algorithms to Achieve Risk Control

We introduce a framework for calibrating machine learning models so that their predictions satisfy explicit, finite-sample statistical guarantees. Our calibration algorithm works with any underlying model and (unknown) data-generating distribution and does not require model refitting. The framework addresses, among other examples, false discovery rate control in multi-label classification, intersection-over-union control in instance segmentation, and the simultaneous control of the type-1 error of outlier detection and confidence set coverage in classification or regression. Our main insight is to reframe the risk-control problem as multiple hypothesis testing, enabling techniques and mathematical arguments different from those in the previous literature. We use our framework to provide new calibration methods for several core machine learning tasks with detailed worked examples in computer vision and tabular medical data.

Poster session #1.

Sensitivity Analysis of Individual Treatment Effects: A Robust Conformal Inference Approach

We propose a model-free framework for sensitivity analysis of individual treatment effects (ITEs), building upon ideas from conformal inference. For any unit, our procedure reports the Γ-value, a number which quantifies the minimum strength of confounding needed to explain away the evidence for ITE. Our approach rests on the reliable predictive inference of counterfactuals and ITEs in situations where the training data is confounded. Under the marginal sensitivity model of Tan (2006), we characterize the shift between the distribution of the observations and that of the counterfactuals. We first develop a general method for predictive inference of test samples from a shifted distribution; we then leverage this to construct covariate-dependent prediction sets for counterfactuals. No matter the value of the shift, these prediction sets (resp. approximately) achieve marginal coverage if the propensity score is known exactly (resp. estimated). We describe a distinct procedure also attaining coverage, however, conditional on the training data. In the latter case, we prove a sharpness result showing that for certain classes of prediction problems, the prediction intervals cannot possibly be tightened. We verify the validity and performance of the new methods via simulation studies and apply them to analyze real datasets.

This is joint work with Ying Jin and Emmanuel Candès.

We develop a general distribution-free framework based on conformal prediction for time series, including prediction intervals for real-valued responses and prediction sets for categorical responses. We show that our intervals and sets asymptotically attain valid conditional and marginal coverages for a broad class of prediction functions and time series. We also show that our interval width or set size converges to the oracle prediction interval or set asymptotically. Moreover, we introduce computationally efficient algorithms called \verb|EnbPI| for prediction intervals and \verb|ERAPS| for prediction sets, which wrap up around ensemble predictors. Our framework is closely related to conformal prediction (CP) but does not require data exchangeability. Both algorithms avoid data-splitting and are computationally efficient by avoiding retraining, thus being scalable to sequentially producing prediction intervals or sets. We perform extensive simulation and real-data analyses to demonstrate their effectiveness compared with existing methods. This is a joint work with Chen Xu at Georgia Tech.

Emmanuel Candès, Victor Chernozhukov, Pietro Perona

Moderator: Stephen Bates

Probabilistic Conformal Prediction Using Conditional Random Samples by Zhendong Wang, Ruijiang Gao, Mingzhang Yin, Mingyuan Zhou and David Blei

On the Utility of Prediction Sets in Human-AI Teams by Varun Babbar, Umang Bhatt and Adrian Weller

Adaptive Conformal Predictions for Time Series by Margaux Zaffran, Olivier Féron, Yannig Goude, Julie Josse and Aymeric Dieuleveut

Approximate Conditional Coverage via Neural Model Approximations by Allen Schmaltz and Danielle Rasooly

Practical Adversarial Multivalid Conformal Prediction by Osbert Bastani, Varun Gupta, Christopher Jung, Georgy Noarov, Ramya Ramalingam and Aaron Roth

VaR-Control: Bounding the Probability of High-Loss Predictions by Jake Snell, Thomas Zollo and Richard Zemel

Confident Adaptive Language Modeling by Tal Schuster and Adam Fisch

Reliable uncertainty quantification is crucial for applying machine learning in safety critical systems such as in healthcare and autonomous vehicles, since it enables decision-making to account for risk. One effective strategy is to construct prediction sets, which modifies a model to output sets of labels rather than individual labels. In this talk, we describe our work on prediction sets that come with probably approximately correct (PAC) guarantees. First, we propose an algorithm for constructing prediction sets that come with PAC guarantees in the i.i.d. setting. Then, we show how our algorithm and its guarantees can be adapted to the covariate shift setting (which is precisely when reliable uncertainty quantification can be most critical). Furthermore, we describe how to adapt our algorithm to the meta learning setting, where a model is adapted to novel tasks with just a handful of examples. Finally, we demonstrate the practical value of PAC prediction sets in a variety of applications, including object classification, detection, and tracking, anomaly detection, natural language entity prediction, detecting oxygen saturation false alarms in pediatric intensive care units, and heart attack prediction.

Conformal prediction is a popular, modern technique for providing valid predictive inference for arbitrary machine learning models. Its validity relies on the assumptions of exchangeability of the data, and symmetry of the given model fitting algorithm as a function of the data. However, exchangeability is often violated when predictive models are deployed in practice. For example, if the data distribution drifts over time, then the data points are no longer exchangeable; moreover, in such settings, we might want to use an algorithm that treats recent observations as more relevant, which would violate the assumption that data points are treated symmetrically. This paper proposes new methodology to deal with both aspects: we use weighted quantiles to introduce robustness against distribution drift, and design a new technique to allow for algorithms that do not treat data points symmetrically, with theoretical results verifying coverage guarantees that are robust to violations of exchangeability. This work is joint with Emmanuel Candes, Aaditya Ramdas, and Ryan Tibshirani.