Modeling Multiattribute Food Decisions Using Semantic Representations (DSI-SRP)
This DSI-SRP fellowship funded Ke Lai to work in the laboratory of Professor Jennifer Trueblood in the Department of Psychology during the summer of 2022. Ke is a senior with majors in Psychology and Economics.
Interested in how people people make decisions or judgments about food items, Ke and Professor Trueblood began by focusing on finding a representation for food items which could be used to model food judgment and decision making using computational tools in natural language processing. They web-scraped recipes from Allrecipe.com, which gave them a sample of about 2000 commonly eaten food items, then used two transformer models, GPT3 and Roberta, to obtain representations for food items. These representations were high dimensional representations of food items.
Ke and Professor Trueblood visualized these representations using dimensionality reduction techniques such as tSNE and PCA. tSNE allowed them to observe that food that belonged to the same category often appeared near each other. PCA allowed them to observe the dimensions that the representations vary on the most. Even though some dimensions could not be interpreted, they observed some important dimensions such as western vs. eastern, taste (salty vs. sweetness), and meatiness of ingredients (salad vs. pork / chicken). Then, they examined the ability of these representations to estimate the similarity between food items by calculating the cosine distance between two vectors.
Going forward, they plan to use these representations to predict human judgments on food items and to collect human judgments on these food items. As a part of psychological science honors project, Ke plans to use a model that uses the vector representations as inputs and predicts human judgement scores. She is also interested in understand how the presence of one food item might change the judgment on another food items. To this end, they have coded this experiment using PsychoPy.
“Before participating in DSI-SRP, I was really a neophyte to computer sciences. Now, I am really glad to see how my skills have been built up through this summer,” Ke reflected. “I have got the hang of web-scraping data using by writing python code that uses Selenium. I have grasped how to use cutting edge tools in artificial intelligence using Word2Vec, GPT-3 and Roberta. I have also learnt to visualize and interpret these high dimensional representations. Finally, I have learned to design experiments to obtain human data using software such as PsychoPy. Overall, this experience has placed me in a good position to use Data Science approaches for my Honors project in the psychology department.”
In addition to receiving support through a DSI-SRP fellowship, this project was supported and facilitated by the DSI Data Science Team through their regular summer workshops and demo sessions.