Reliability Evaluation of Individual Predictions: A Data-centric Approach

TL;DR

This paper introduces data-centric reliability measures for individual predictions in machine learning, focusing on training data similarity and local uncertainty to assess prediction trustworthiness, with scalable algorithms and extensive validation.

Contribution

It proposes novel, scalable algorithms for data-centric reliability assessment of individual predictions, independent of model certainty, applicable to large and high-dimensional datasets.

Findings

Reliability measures correlate with model performance across datasets.

Algorithms are efficient and scalable for large, multi-dimensional data.

Extensive experiments validate the effectiveness of the proposed measures.

Abstract

Machine learning models only provide probabilistic guarantees on the expected loss of random samples from the distribution represented by their training data. As a result, a model with high accuracy, may or may not be reliable for predicting an individual query point. To address this issue, XAI aims to provide explanations of individual predictions, while approaches such as conformal predictions, probabilistic predictions, and prediction intervals count on the model's certainty in its prediction to identify unreliable cases. Conversely, instead of relying on the model itself, we look for insights in the training data. That is, following the fact a model's performance is limited to the data it has been trained on, we ask "is a model trained on a given data set, fit for making a specific prediction?". Specifically, we argue that a model's prediction is not reliable if (i) there were not enough similar instances in the training set to the query point, and (ii) if there is a high fluctuation (uncertainty) in the vicinity of the query point in the training set. Using these two observations, we propose data-centric reliability measures for individual predictions and develop novel algorithms for efficient and effective computation of the reliability measures during inference time. The proposed algorithms learn the necessary components of the measures from the data itself and are sublinear, which makes them scalable to very large and multi-dimensional settings. Furthermore, an estimator is designed to enable no-data access during the inference time. We conduct extensive experiments using multiple real and synthetic data sets and different tasks, which reflect a consistent correlation between distrust values and model performance.