Abstract:
|
While many machine learning (ML) models have strong predictive value, they are usually considered a “black box” where the contributing role of each individual variable is unknown. This uncertainty leads to serious challenges in identifying causal relations in ML models and is a critical barrier to the development and use of ML. We have developed multivariate versions of the e-value, using the foundation of mixture modeling, to identify the degree of causality of both individual variables and individual clusters. We also use these metrics in tandem with estimates of standard error and entropy in mixture models for further assessing bias and error in unsupervised ML approaches. In this presentation, we discuss these new approaches and demonstrate the usefulness of the methodology.
|