Computational appraisal of gender representativeness in popular movies

Traditional research on gender representation in mass media, including movies, has relied heavily on qualitative content analysis, which is time-consuming and difficult to scale. This paper argues for the application of computational methods to overcome these limitations. The authors aim to demonstrate how automated methods can enhance the scope and resolution of gender representation studies, providing a more comprehensive and nuanced understanding of trends and patterns. This approach allows for large-scale analysis across temporal and other variables, providing insights not easily accessible through manual methods. The study focuses on popular movies, defining popularity not solely based on box office success, but also considering the activity and engagement on online platforms like peer-to-peer file-sharing networks and IMDb. This broader definition aims to capture a wider range of mainstream movies.

Existing research on gender representation in media, dating back to Busby (1975), has consistently highlighted the underrepresentation and sexualization of women across various media. While some studies have utilized large-scale datasets from sources like IMDb, they are limited by the availability of metadata. Recent advancements in AI and data science offer new possibilities for automated analysis of text, image, and video content, including face recognition, scene analysis, and narrative extraction. However, these techniques have been mainly applied within their originating scientific fields, rather than social science research. This study bridges this gap, employing AI to analyze gender representation in a large-scale dataset of movies.

The study compiled a dataset of 3776 popular movies from 1985 to 2019, selected based on data from a peer-to-peer file-sharing network (YIFY) and IMDb. Frames were extracted from these movies at a frequency of one image every 2 seconds, resulting in over 12.4 million images. Wolfram Mathematica's face detection and gender estimation algorithms were used to analyze these images. To address potential biases in the algorithms, a human evaluation protocol was implemented. 1000 images (500 female and 500 male faces) were manually reviewed to assess the accuracy of face detection and gender inference. The algorithm exhibited 92% accuracy for face detection and 73.9% accuracy for gender inference, with a bias towards misclassifying faces as female. These error rates were then used to generate correction factors for the female face ratio (FFR) calculations. The FFR, defined as the percentage of faces identified as female in a movie, was the primary metric used to quantify gender representation. This was compared with the Bechdel test results, which were manually compiled for a subset of the data. Additionally, the study explored the spatial positioning of faces (mise-en-cadre) to analyze the way characters were framed on screen. This analysis utilized a 3x3 grid system to categorize face position within a frame.

The overall average FFR across all movies was 34.52%, comparable to findings in existing literature. However, the FFR varied significantly across genres. A key finding was a clear temporal trend towards increased female representation. The FFR increased significantly from an average of 27% in the period 1985-1998 to 44.9% in 2014-2019, approaching gender parity. This trend was also observed in the percentage of movies passing the Bechdel test. The analysis also revealed correlations between the FFR and audience-related factors such as movie budget, gross revenue, and user ratings, suggesting a preference for movies with FFRs closer to the overall average (and thus exhibiting underrepresentation of women). The study of face-ism showed minimal differences between male and female characters in terms of the proportion of the frame occupied by their faces, although this analysis was limited by the inability of the algorithm to account for body size. Finally, the analysis of mise-en-cadre revealed a statistically significant difference in the vertical position of male and female faces, particularly in gender-mixed scenes, with women more frequently appearing in the middle third of the frame and men in the upper third. This bias was partly explained by height differences between male and female actors.

The findings challenge the common assertion in existing literature that gender representation in movies has remained relatively static. The study's methodology, employing a broader definition of 'popularity' and a large-scale dataset, revealed a substantial trend toward increased female on-screen presence. The correlation between FFR and the Bechdel test, along with the analysis of audience-related features, indicates that while female on-screen presence may be improving, this doesn't necessarily translate into a fairer or more nuanced representation of women in terms of narrative roles and impact. The observed spatial bias in face placement suggests subtle yet significant differences in how male and female characters are visually framed, influencing viewer perception. These results highlight the potential of combining quantitative and qualitative approaches to better understand the complexities of gender representation in movies.

This study demonstrates the value of computational methods for analyzing gender representation in mass media, providing a large-scale, high-resolution view of historical trends. While the study shows progress toward greater female on-screen presence, it also reveals the importance of more sophisticated analyses to assess the depth and nuance of gender representation, beyond simply the numerical count of female faces. Future research could integrate qualitative methods, focus on the analysis of body placement along with facial features and examine the relationship between on-screen representation and audience engagement more deeply.

The accuracy of the gender classification algorithm, while relatively high, introduced some bias, necessitating correction factors. Furthermore, the definition of popularity used in movie selection may not perfectly capture all aspects of mainstream cinematic culture. Finally, the analysis of face-ism and mise-en-cadre was limited by the algorithm's inability to accurately assess body size and other contextual factors that might influence character framing and their implied dominance.