researchers have created an algorithm for automatic prediction of personality traits from the “Big five” on a photograph of a human face. To train the algorithm and test it the scientists collected more than 30 thousand images 12,4 thousand men and women. The correlation coefficient between predicted and actual feature is a positive but quite small (equal to 0,243), although the probability of determining the correct character (approximately 58 percent) is still above the occasional hit. Article published in the journal Scientific Reports.
In the process of communication a person makes conclusions about the interlocutor not only in conversation; also important non-verbal means of communication, particularly emotions, facial expression and overall appearance. Studies show that different features of the person’s appearance can determine, for example, is sick whether he is capable of if he’s cheating.
While it is clear that the definition of some peculiarities of conduct or character in appearance requires the system, there are certain external features should always be associated with certain personality traits. From this dependence there is some biological basis: on the one hand, be inherited can and traits, and external features, which means that the relationship between them in fact can last for generations. On the other hand, and on the traits and external traits affect prenatal development: for example, elevated levels of testosterone.
One of the most effective ways of assessing such dependency — check automatically on a large dataset: a similar method three years ago, scientists were able to learn how to determine the person’s sexual orientation. Alexander Kachur (Alexander Kachur) of the company AIPictor and his colleagues to assess the relationship between facial features and character gathered 31367 photos 12447 men and women. All photos were taken in good lighting and full face, and participants were asked not to apply makeup and not to wear jewelry, and keep a neutral facial expression. In addition, all study participants filled in a standardized questionnaire on the defining traits of the “Big five”: extraversion, friendliness, conscientiousness, neuroticism and openness to experience.
The resulting photographs then marked out using residual convolutional neural networks ResNet so that you can highlight certain facial features, for which people can be classified. These features are then correlated with the results of the questionnaire “Big five” using a multilayer perceptron, having the correlation coefficient between the predicted trait and the present.
The correlation coefficient for all traits, certain photograph of the person was positive, but varied strongly: the worst predicted openness to experience women (the coefficient of 0,137), and best of all — honesty, and for both men and women (coefficients and 0,358 0,386 respectively).