CIML-R: Causally Informed Machine Learning Based on Feature Relevance

Applications relying on machine learning and statistical learning techniques, such as neural networks, have significantly grown over the past decade. Nevertheless, as these techniques learn only from observational data, they suffer from spurious correlations that may limit their performance in domai...

Full description

Saved in:
Bibliographic Details
Main Authors: Surner, Martin, Khelil, Abdelmajid
Format: Conference Proceeding
Language:English
Subjects:
Accuracy
Causality
Correlation
Domain Generalization
Informed Machine Learning
Knowledge engineering
Neural networks
Production
Statistical learning
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Applications relying on machine learning and statistical learning techniques, such as neural networks, have significantly grown over the past decade. Nevertheless, as these techniques learn only from observational data, they suffer from spurious correlations that may limit their performance in domain shifts. In this paper, we address this issue. We propose an approach that guides neural networks during the training phase using explanations. User-defined causal knowledge is incorporated to guide the considered neural network. Such guidance reduces the risk of spurious correlations, allowing for a focus on the causal structure. Our method extends the feed-forward step of neural networks to track the feature relevance. The resulting feature relevance is accounted for in a regularization term in the loss function. We compare the performance of a conventional neural network with and without the proposed extension. This comparison indicates that the learning process successfully adopts the causal mechanism, while the conventional approach learns based on spurious correlations.
ISSN:2835-3420
DOI:10.1109/SDS60720.2024.00018