Is Beauty in the Eye of the Biased Machine?

A group of M.S. in Data Science students at the University of Virginia used their final project in an introductory machine learning course to explore the Golden Ratio — a mathematical symbol of perfection that has been recently used to define beauty standards.

Associate professor Prince Afriye, who teaches the Machine Learning I: Introduction to Predictive Modeling course, said he intentionally kept the rubric for the project open so students could choose datasets that genuinely interested them and build models to generate meaningful insights.  

For this group, the topic of interest bridged data science with modern standards of beauty. We spoke with students from Afriye’s class to learn more about the inspiration for their project and what their research illuminated.  

Q: There is a social media trend where people photoshop celebrity faces to be proportionate with the Golden Ratio. What inspired the focus of your project and what did you want to communicate through its completion?  

Stephanie Delgadillo-Cartagena: The inspiration for this project came from the broader idea that beauty is sometimes treated as something that can be minimized into measurable patterns and proportions. The Golden Ratio is referred to a lot in art and nature as a symbol of balance or harmony, and we were interested in exploring how that concept translates into facial aesthetics. 

Through the completion of this project, we wanted to communicate how mathematical models and deep learning tools can be applied to real world questions about beauty. However, at the same time we also wanted to highlight the limitations of trying to reduce something as subjective as beauty to numerical proportions. 

Q: The Golden Ratio has been fashioned into an unobtainable standard of beauty, and its origins have seemingly been forgotten. Did these biases impact your results?  

Jillian Howe: The Golden Ratio has increasingly been framed as an objective and universal standard of beauty, while its historical and mathematical origins are often overlooked. Rooted in the mathematical constant φ (phi), approximately 1.618, the Golden Ratio describes a proportional relationship derived from geometric principles and observed in patterns such as the Fibonacci sequence and certain natural phenomena. 

In our study, we applied computational models to examine whether facial proportions in our dataset exhibited patterns consistent with this ratio, and whether such patterns varied across demographic groups. Our findings indicate that correlations between facial proportions and demographic characteristics were more pronounced than adherence to the Golden Ratio itself. This suggests that what is often labeled as a “natural” or universal standard may instead reflect underlying demographic variation. Consequently, the perception of the Golden Ratio as an objective benchmark of beauty may be influenced by inherent population differences rather than a single, universally applicable proportion.

Q: Tell us about the process of training your model.  

Bella Lu: To create our dataset, we collected professional headshots from 52 members of our cohort. To make the project engaging and personal, we invited our classmates to submit their headshots through a Microsoft form along with demographic information like gender and race. (We baked and distributed cookies as an incentive!) 

We used multiple methods and packages to be able to analyze the images but finally settled on Google’s MediaPipe to perform facial feature detection and image analysis. From each photo, we extracted measurements such as face width, face height, nose width, mouth width, pupil distance, and related ratios like mouth-to-nose ratio. These features were combined to compare to a “golden ratio” score, forming the complete dataset used for modeling.  

With our custom dataset, we implemented five models as a requirement for our project: linear regression, logistic regression, KNN, K-means clustering, and PCA. Throughout the process, we encountered several obstacles. First, our sample size was small and not representative of a broader population. Second, not everyone sent in their professional headshot; some people sent in selfies with lower resolution. The variation in posture, lighting, and image quality affected feature extraction, and facial detection required significant trial and error. Early models misidentified features, detected background elements as faces, or failed to detect any face at all. After testing multiple convolutional neural network approaches, we selected the one that provided the most consistent and accurate facial measurements for our analysis.  

Q: Why do you think we assign meaning to numbers and mathematic equations?  

Sheyi Faparusi: I think we assign meaning to numbers and mathematic equations because the world is so vast, and there are so many things that are incomprehensible and hard to quantify. When you can quantify something, write a formula for it, or spot a pattern, it can make things simpler and sometimes more interesting. 

I think in some ways the numbers are neutral, but they are not neutral in the way they are utilized and also for the reasons people search for them. For example, the golden ratio as a metric for beauty is a biased metric. It doesn’t work or mean anything for all groups, and beauty is not something that is so easily quantified.  

Q: During your presentation, you mentioned that beauty cannot be defined. How can models with pre-existing human bias be trained to function with this worldview?  

Sophie Kim: In all honesty, I don’t think a model with any sort of pre-existing human bias could be trained to function accurately, especially when analyzing something like human beauty.  Beauty is a really tough thing to define objectively and analyze/predict accurately using a model.  

Though, I think there could be some sort of robust model that could get somewhere close to this. The key idea in that case would be to have the model learn that beauty is really not something that can be classified in a Yes/No sort of way. The training data would have to be extremely diverse so that no defining beauty standard could dominate. And I don’t think the end goal of the model would be to say, “Face A is more beautiful than Face B.” I think a more accurate (and ethical) end goal of the model would be “Face A is beautiful in X way, and Face B is beautiful in Y way.”  

But at the end of the day, any sort of model that tries to analyze human beauty simply can’t be trusted. After all, what could a computational model really say about what the human eye truly sees? Beauty is not about the feature variables of the face, but rather, what the human observer sees in that face. I think that’s an important gap in translation that a model could not be trained on.