Confounder-Invariant Representation Learning (CIRL) for Robust Olfaction with Scarce Aroma Sensor Data: Mitigating...

Area of Science:

• Artificial intelligence
• Sensory science
• Machine learning

Background:

• Olfactory aroma data is susceptible to confounding factors like high humidity, which obscure volatile organic compound (VOC) patterns.
• Traditional representation learning requires extensive datasets to manage confounder variance, posing challenges for data-scarce applications.

Purpose of the Study:

• To introduce Confounder-Invariant Representation Learning (CIRL), a novel method for mitigating confounding influences in limited-data settings.
• To enhance learned representations by explicitly using confounder information, such as relative humidity, to improve data purity and model robustness.

Main Methods:

• CIRL was developed to leverage explicit confounder information, like relative humidity, to reduce environmental effects on sensor data.
• CIRL was integrated with standard autoencoder models and applied to three distinct breath aroma datasets (acetone, ketosis, peppermint-oil breath).

Main Results:

• CIRL demonstrated improved generalization performance, increasing classification accuracy by 10-15% across all tested breath aroma datasets.
• The method effectively reduced the impact of high humidity, a significant confounding factor, on the sensor data.

Conclusions:

• CIRL shows significant potential for improving the reliability of artificial olfaction systems, particularly in real-world conditions with confounding variables.
• The findings suggest CIRL can advance breath-based diagnostics and other applications requiring robust olfactory data analysis.