US 11,636,383 B2 · Detecting and preventing unwanted model training data

Abstract

A method and a system for dataset evaluation are provided. The method can include acquiring a new dataset for integration with an existing dataset. The existing dataset is used to train a machine learning model. The method can also include calculating a baseline of variation for the existing dataset. The baseline of variation can be determined by evaluating various characteristics of the existing dataset. The method can further include determining an output value for the new dataset by also evaluating characteristics of the new dataset. The method can also include comparing the output value to the baseline of variation to generate a variance between the output value and the baseline of variation. The method can further include determining whether the variance is within an acceptable range of the baseline of variation.

Claims18

Dependent claims

A system comprising:

a data pre-processor configured to perform data analytics on a new dataset and compare the data analytics from the new dataset with analytics from an existing dataset;
a machine learning model configured to be trained from the existing dataset;
a model tester configured to test behaviors performed by the machine learning model, and to compare a set of behaviors performed by the machine learning model prior to being trained with the new dataset with the set of behaviors performed by the machine learning model subsequent to being trained with the new dataset;
and
a fairness analyzer configured to test predictions made by the machine learning model for biases and measure the biases in the test predictions, wherein the biases include automation bias detectable in the test predictions.

2 more elements

A computer program product for dataset evaluation, the computer program product comprising a computer readable storage medium having program instruction embodied therewith, the program instructions executable by a processor to cause the processor to:

acquire a new dataset for integration with an existing dataset, wherein the existing dataset is used to train a model;
calculate a baseline of variation for the existing dataset by evaluating characteristics of the existing dataset;
determine an output value by evaluating characteristics of the new dataset;
compare the output value to the baseline of variation to generate a variance of the output value to the baseline of variation;
determine whether the variance is within an acceptable range of the baseline of variation;
and
transforming, upon determining that the variance is within the acceptable range of the baseline of variation, the existing dataset by adding the new dataset to the existing dataset.

4 more elements

Dependencies are inferred from the claim text.

Detecting and preventing unwanted model training data

Life of this patent

Detecting and preventing unwanted model training data

Life of this patent

Abstract

Claims18