Assess the Big Five personality traits using static facial images from real life.

Key Chapter Title List

1. Research Background and Theoretical Foundation
2. Research Objectives and Hypotheses
3. Sample and Research Procedure
4. Ethical Approval
5. Data Screening
6. Big Five Personality Measurement Tools
7. Image Screening and Preprocessing
8. Neural Network Architecture
9. Research Results
10. Discussion
11. Research Limitations
12. Data Availability Statement

Document Introduction

Extensive research has confirmed that the morphological features and social cues of the human face can convey signals related to personality and behavior. Although previous studies have explored the association between artificially synthesized facial images and personality trait attribution, systematic prediction of all Big Five personality traits for both genders based on static facial images from real life remains to be improved. Existing studies also suffer from issues such as small sample sizes, significant methodological differences, and insufficient consistency in results. The core aim of this study is to fill this gap and validate the feasibility of extracting personality cues from static facial images using machine learning algorithms.

The study utilized 31,367 real-life facial images provided by 12,447 anonymous volunteers, all of whom completed self-report measures of the Big Five personality traits. To ensure data quality, the study excluded invalid questionnaires, low-quality images, and fabricated content through multi-stage screening. The final dataset was divided into training and test sets in a 9:1 ratio, used for model training and validation, respectively. Considering the sexual dimorphism of facial features and some personality traits, all prediction models were trained and validated separately for males and females.

The study employed a two-layer machine learning algorithm architecture: first, a computer vision neural network (NNCV) was constructed based on the ResNet architecture to extract a 128-dimensional invariant feature vector from static facial images; subsequently, a personality diagnostic neural network (NNPD) was trained to predict Big Five personality trait scores using this feature vector as input. The data processing strictly adhered to the Declaration of Helsinki, was approved by the Research Ethics Committee of the Open University of Humanities and Economics, and all participants provided informed consent.

Key findings of the study show statistically significant correlations between the artificial neural network's predictions of Big Five personality traits and self-reported scores, with an average effect size of 0.243, exceeding the results of previous studies using selfie images. Among them, the prediction correlation for Conscientiousness was the highest (male 0.360, female 0.335), while the prediction accuracy for Openness was the lowest. The prediction effectiveness for Extraversion and Neuroticism traits was significantly better for females than for males. Inter-trait correlation analysis indicated that the correlation structure of the predicted scores partially differed from that of the self-report scales, suggesting that the General Factor of Personality (GFP) may have a biological basis.

This study confirms that even real-life images taken under uncontrolled conditions can effectively predict personality traits through complex computer vision algorithms. Its methodological advantage lies in not relying on 3D scans or high-precision facial markers, requiring only an ordinary desktop computer. The results not only provide new empirical support for the face-personality association but also offer potential application scenarios for rapid personality assessment, which can assist in areas such as product-service matching, interpersonal interaction pairing, and personalization of human-computer interaction.

It should be noted that the sample of this study primarily consists of Russian-speaking Caucasian adults, with limitations in geographical and cultural scope. Furthermore, additional cues present in real facial images, such as makeup and camera angles, may influence the prediction results. Future research needs to expand sample diversity and further distinguish the roles of facial morphological features from other image cues.