Intuitive Data-Driven Visualization of Food Relatedness via t-Distributed Stochastic Neighbor Embedding

Abstract

The relationship between diet and health is important, yet difficult
to study in practice. Dietary pattern analysis is one method for
investigating this link; having more variety in diet tends to be bene-
ficial and a score can be generated based on a heuristic approach
to food intake habits. We aim to enhance the intuition behind
these food scores by creating an intuitive data-driven visualization
of food relatedness by leveraging t-distributed stochastic neighbor
embedding (t-SNE). More specifically, by performing t-SNE anal-
ysis in a controlled manner to project the high-dimensional nutri-
tional information of food items into a lower dimensional food sim-
ilarity space, the natural clustering of foods based on the underly-
ing nutritional composition becomes visually observable. The effi-
cacy of this data-driven approach for visualizing food relatedness
was investigated on a total of 8549 food item entries in the USDA
food composition database, with the results showing considerable
promise as a tool for gaining important nutritional insights. This is
the first step toward providing a novel method to enhance dietary
pattern analysis with additional context and insight into food intake
habits based on the inherent nutritional content of the foods con-
sumed.