Gaussian Mixture Models for Identifying Ultra-Processed Foods

Ultra-processed foods (UPFs) are industrial formulations of multiple ingredients to make a highly palatable food product and account for nearly half of the average British diet and most developed countries. UPFs are strongly associated with obesity, chronic disease, and increased mortality. However, the identification of UPFs is currently still heavily relied on by human experts, making consistent and large-scale classification very difficult and resource intensive. This research explores a semi-supervised pipeline in which available labels are leveraged during preprocessing and performance evaluation, but not used in unsupervised clustering tasks. We investigated and tested Gaussian Mixture Models (GMM) and few other commonly used clustering methods on real-world data, and then compared them with a baseline method k-means. This reflects the real-world scenario where food databases contain partially labelled data but full annotation remains impractical. These results demonstrate that unsupervised learning offers a viable and scalable option for automated food categorization where labelled data is scarce.