Many people can guess the personality traits of the interlocutor by facial features. Experts in non-verbal communication do this even from a photograph. Is it possible to teach artificial intelligence like this – in a study by the National Research University Higher School of Economics
The material is provided by the Trends portal IQ.HSE.RU.
Experts in psychology and artificial intelligence from the Higher School of Economics, BestFitMe (UK and our country) and Artificial Intelligence LLC (our country) taught a cascade neural network to identify personality traits from the Big Five from a photograph of a face. The study was published in the journal Scientific Reports. The test dataset is available in the repository.
The big five personality traits include:
What are we talking about?
Physiognomy – the definition of a person’s character by his appearance – has long been popular in European culture. Having originated in the writings of the ancient Greek thinkers Aristotle and Theophrastus, it reached its maximum development in the XVIII-XIX centuries in the works of Johann Lavater, Charles Darwin and Cesare Lombroso. As a result of the empirical verification of the ideas of these authors in the XNUMXth century, physiognomy was recognized as a pseudoscience. However, recent studies show that the relationship between facial features and personality traits still exists.
First, it has been established that a number of psychological characteristics are genetically predetermined. For personality traits from the “Big Five”, the contribution of the hereditary factor varies from 30 to 60%. Genetics also determines the shape of the bones of the skull, which sets individual facial features. It is believed that the evolution and relationship of characterological properties was influenced by sexual selection. Women were looking for clear signs of useful or safe character traits, so they acquired sexual attractiveness (for example, positive selection for forehead height), and after it, heritability.
In addition, the shape of the face and behavior is influenced by prenatal and postnatal exposure to hormones. The outline and size of the cheekbones, lower jaw, the ratio of the length and width of the face, as well as a number of other features are a visible indicator of the level of sex hormones – testosterone and estrogen. They are also responsible for the propensity to take risks, aggressiveness, the desire for competition and dominance, or, conversely, gentleness and compliance, tenderness, caring.
A number of studies show the relationship of facial features with a predisposition to certain behaviors, expressed as five basic personality traits.
Usually people quite accurately guess individual features of a character by facial features. For professional psychologists and non-verbal communication specialists, the correlation between their predictions and data from personality questionnaires is even higher. And this means that artificial intelligence can also be trained in a similar skill if a dataset of sufficient size and well pre-labeled for training is created.
How did you study?
In their new work, specialists from the Higher School of Economics and technology startups have collected 77 face photographs from 346 volunteers through social networks. All photos were taken using webcams under controlled conditions (neutral facial expression, frontal view, looking into the lens, good lighting and no makeup or jewelry). Respondents were also asked to take an improved version of the online 25PFQ questionnaire to determine their personality profile and the severity of the Big Five psychological traits.
The final dataset contained 12 descriptions of personality traits from the questionnaires and 447 photos. The dataset was randomly divided into two parts. The first one (31% of the data) served as a training sample for the neural network. The second (367%) is a control for assessing the predictive capabilities of the trained algorithm.
Initially, the neural network was trained to distinguish the faces of different people, but to consistently recognize the face of the same person. Then the algorithm was trained to decompose each image into 128 invariant features – regularly repeating individual features. Within the model, each invariant was represented as a vector in 128-dimensional space.
The data thus obtained entered the multilayer perceptron. In it, artificial neurons compared signs with the personal characteristics of volunteers. If they matched, the data was “fixed”. And if there was a discrepancy, then the calculated error was again fed to the input of the neural network. Gradually, artificial intelligence learned to more accurately compare facial features and character.
What did you get?
The first published results are still far from ideal and rather look like a proof of concept. The correlation coefficient between questionnaire data and algorithm predictions ranged from a small 0,14 to a reassuring 0,36. The average effect size, a statistic of the model’s practical relevance, was 0,24. It follows from this that the algorithm made the correct conclusion almost 60% of the time, while random guessing usually matches only 50%. The superiority of 10% seems insignificant, but in fact, in terms of accuracy of predictions, artificial intelligence is significantly ahead of people if they judge by the facial features of a stranger.
What does it do?
With further improvement in the quality of the algorithm, it can be used in recommender systems of online stores and online services. Great prospects are also opening up for HR departments – quick psychological diagnostics right during a zoom interview with job candidates. The method will be especially effective in the case of mass recruitment of personnel in the service sector – taxi drivers, sellers, cleaners, etc. With its help, you can quickly weed out aggressive, mentally unstable and unscrupulous people.
Another possible market is dating apps and sites, as well as services for the psychological assessment of liked strangers from social networks. Their use can seriously increase the safety of women when meeting in person with men who know only through online correspondence.
Subscribe to the Trends Telegram channel and stay up to date with current trends and forecasts about the future of technology, economics, education and innovation.