Deep neural networks can quantify facial characteristics more accurately than previous methods, improving predictions of in-person attraction, according to a study published in Evolution & Human Behavior.
Facial attractiveness significantly influences human mate selection, yet current methods for measuring key facial traits like masculinity, similarity, and averageness have important limitations. Traditional approaches have relied on either subjective ratings, which are vulnerable to biases, or objective landmark-based measurements that map geometric facial structures using predefined points. While the latter offers quantitative data, these methods fail to capture important visual details such as skin texture, hair color, and facial contrast.
To overcome these limitations, Amy A.Z. Zhao and Brendan P. Zietsch investigated the potential of deep neural networks as an alternative approach. These machine-learning techniques offer a more comprehensive method for quantifying facial characteristics that addresses the shortcomings of both subjective and landmark-based assessments.
Research suggests masculinity in male faces often correlates with attractiveness, while facial similarity can signal trustworthiness or kindness. However, these findings have typically emerged from studies using static facial images rather than real-life interactions.
Zhao and Zietsch’s work explores whether deep neural networks can generate facial metrics that not only improve measurement accuracy but also better predict attraction in naturalistic settings.
The study analyzed data from 682 participants (344 women) recruited from the University of Queensland, who took part in a speed-dating experiment. To ensure consistency, all participants were required to be single, heterosexual, and fluent in English. The researchers obtained in-person ratings of facial attractiveness and kindness from 2,285 speed-dating interactions, in which participants rotated through short conversations with potential romantic partners.
To measure facial traits, the researchers collected standardized facial images of all participants under consistent lighting conditions with a fixed camera setup. They then applied three different methods to quantify facial characteristics: 1) manually placed facial landmarks, where research assistants placed 28 predefined points on participants’ faces; 2) automatically detected facial landmarks using an artificial intelligence-based system that placed 83 points; and 3) deep neural networks that extracted 4,096 facial feature coordinates from each image.
These methods were used to assess three key facial traits—averageness, masculinity, and similarity—by comparing facial features across participants and analyzing how well they predicted in-person ratings of attractiveness and kindness.
The findings showed that facial measures derived from deep neural networks could predict in-person ratings of facial attractiveness and kindness as well as, or better than, traditional landmark-based methods. Notably, neural network-derived masculinity scores for male faces were strongly correlated with attractiveness ratings, reinforcing previous findings that masculine male features are considered appealing. In contrast, masculinity in female faces was negatively correlated with attractiveness, meaning more feminine female faces were rated as more attractive.
Furthermore, neural networks provided a more robust measure of masculinity by avoiding a common limitation of landmark-based methods: their tendency to be influenced by facial pitch (the angle of head tilt in images). When measured using neural networks, masculinity scores were largely unaffected by the angle at which a person held their head in the photograph.
The study also found evidence that individuals with more masculine or feminine faces preferred similarly sex-typical partners, a pattern known as assortative mating. This effect was detected using neural networks but not with traditional landmark-based measurements, suggesting that deep learning methods may be better at capturing subtle facial features that influence real-life attraction.
Additionally, facial similarity was associated with kindness ratings, supporting prior research indicating that people tend to perceive those who look like them as more trustworthy and prosocial. This effect was evident when using both automatic landmarking and neural network-derived similarity scores but was less pronounced when measured through manual landmarks.
Facial averageness—often considered a key feature of attractiveness—was also a significant predictor of positive ratings, particularly when measured using neural networks and automatic landmarks.
One limitation is the lack of transparency in deep neural network models. Unlike landmark-based methods, which provide clear measurements of specific facial features, neural networks operate as “black boxes,” making it difficult to determine which aspects of a face contribute to a given rating.
The research, “Deep neural networks generate facial metrics that overcome limitations of previous methods and predict in-person attraction,” was authored by Amy A.Z. Zhao and Brendan P. Zietsch.
