Iteratively trained machine learning models for evaluations of internal consistency

Abstract

Disclosed is an approach for evaluating internal consistency of object classifications using machine learning modeling. In an example, models are iteratively trained, using supervised learning, on different majority segments (e.g., about 90%) of a dataset as training data segments. Trained models can be applied to the remaining data (e.g., about 10%) as test data segments to obtain, for each object, a predicted classification and a confidence score. Models in training iterations (e.g., 10 iterations) may be independently trained on substantially non-overlapping test data segments (with each iteration testing, e.g., about 10% of the dataset). When a model's predicted classification disagrees from a prior classification, and the confidence of the prediction is sufficiently high (indicating sufficiently strong disagreement), the object's prior classification may be revised. Training data, other than the data itself being evaluated for consistency, is not necessarily required.