Abstract Model monitoring is a critical component of the machine learning lifecycle, safeguarding against undetected drops in the model’s performance after deployment. Traditionally, performance monitoring has required access to ground truth labels, which are not always readily available. This can result in unacceptable latency or render performance monitoring altogether impossible. Recently, methods designed to estimate the accuracy of classifier models without access to labels have shown promising results. However, there are various other metrics that might be more suitable for assessing model performance in many cases. Until now, none of these important metrics has received similar interest from the scientific community. In this work, we address this gap by presenting Confidence-based Performance estimation (CBPE), a novel method that can estimate any binary classification metric defined using the confusion matrix. In particular, we choose four metrics from this large family: accuracy, precision, recall, and $$\hbox {F}_1$$ , to demonstrate our method. CBPE treats the elements of the confusion matrix as random variables and leverages calibrated confidence scores of the model to estimate their distributions. The desired metric is then also treated as a random variable, whose full probability distribution can be derived from the estimated confusion matrix. CBPE is shown to produce estimates that come with strong theoretical guarantees and valid confidence intervals.

After a machine learning model has been deployed into production, its predictive performance needs to be monitored. Ideally, such monitoring can be carried out by comparing the model’s predictions against ground truth labels. For this to be possible, the ground truth labels must be available relatively soon after inference. However, there are many use cases where ground truth labels are available only after a significant delay, or in the worst case, not at all. In such cases, directly monitoring the model’s predictive performance is impossible. Recently, novel methods for estimating the predictive performance of a model when ground truth is unavailable have been developed. Many of these methods leverage model confidence or other uncertainty estimates and are experimentally compared against a naive baseline method, namely Average Confidence (AC), which estimates model accuracy as the average of confidence scores for a given set of predictions. However, until now the theoretical properties of the AC method have not been properly explored. In this paper, we bridge this gap by reviewing the AC method and show that under certain general assumptions, it is an unbiased and consistent estimator of model accuracy. We also augment the AC method by deriving valid confidence intervals for the estimates it produces. These contributions elevate AC from an ad-hoc estimator to a principled one, encouraging its use in practice. We complement our theoretical results with empirical experiments, comparing AC against more complex estimators in a monitoring setting under covariate shift. We conduct our experiments using synthetic datasets, which allow for full control over the nature of the shift. Our experiments with binary classifiers show that the AC method is able to beat other estimators in many cases. However, the comparative quality of the different estimators is found to be heavily case-dependent.

After deployment, machine learning models often experience performance degradation due to shifts in data distribution. It is challenging to assess post-deployment performance accurately when labels are missing or delayed. Existing proxy methods, such as data drift detection, fail to measure the effects of these shifts adequately. To address this, we introduce a new method for evaluating binary classification models on unlabeled tabular data that accurately estimates model performance under covariate shift and call it Probabilistic Adaptive Performance Estimation (PAPE). It can be applied to any performance metric defined with elements of the confusion matrix. Crucially, PAPE operates independently of the original model, relying only on its predictions and probability estimates, and does not need any assumptions about the nature of covariate shift, learning directly from data instead. We tested PAPE using over 900 dataset-model combinations from US census data, assessing its performance against several benchmarks through various metrics. Our findings show that PAPE outperforms other methodologies, making it a superior choice for estimating the performance of binary classification models.

After a machine learning model has been deployed into production, its predictive performance needs to be monitored. Ideally, such monitoring can be carried out by comparing the model’s predictions against ground truth labels. For this to be possible, the ground truth labels must be available relatively soon after inference. However, there are many use cases where ground truth labels are available only after a significant delay, or in the worst case, not at all. In such cases, directly monitoring the model’s predictive performance is impossible. Recently, novel methods for estimating the predictive performance of a model when ground truth is unavailable have been developed. Many of these methods leverage model confidence or other uncertainty estimates and are experimentally compared against a naive baseline method, namely Average Confidence (AC), which estimates model accuracy as the average of confidence scores for a given set of predictions. However, until now the theoretical properties of the AC method have not been properly explored. In this paper, we bridge this gap by reviewing the AC method and show that under certain general assumptions, it is an unbiased and consistent estimator of model accuracy. We also augment the AC method by deriving valid confidence intervals for the estimates it produces. These contributions elevate AC from an ad-hoc estimator to a principled one, encouraging its use in practice. We complement our theoretical results with empirical experiments, comparing AC against more complex estimators in a monitoring setting under covariate shift. We conduct our experiments using synthetic datasets, which allow for full control over the nature of the shift. Our experiments with binary classifiers show that the AC method is able to beat other estimators in many cases. However, the comparative quality of the different estimators is found to be heavily case-dependent.