FoodEx2vec: New foods' representation for advanced food data analysis

Area of Science:

• Food science
• Toxicology
• Computational methods

Background:

• Vast amounts of food and toxicology data require advanced analysis.
• Increased avocado consumption necessitates understanding its public health and environmental impact.
• Manual indexing of food data is time-consuming and may miss nuances.

Purpose of the Study:

• To present a novel approach for representing food data using continuous numerical vectors (food embeddings).
• To evaluate the utility of food embeddings in various food data analysis tasks.
• To demonstrate the superiority of food embeddings over traditional methods.

Main Methods:

• Developed a method to convert food items into vectors of continuous numbers (food embeddings).
• Evaluated the approach through four tasks: food group determination, food class detection (raw, derivative, composite), similarity identification, and expert evaluation.
• Compared the performance of food embeddings against traditional data representation methods.

Main Results:

• Food embeddings successfully automated the determination of food groups and food classes.
• The method effectively identified similar food concepts.
• Qualitative evaluation by a food expert supported the utility of the approach.
• Vector representations significantly outperformed traditional methods in food data analysis.

Conclusions:

• Food embeddings offer a powerful alternative to manual indexing for food data.
• This approach represents a significant advancement in food data analysis, enabling new knowledge discovery.
• The method has broad applications in understanding food properties and impacts.