On identification of big-five personality traits through choice of images in a real-world setting

摘要

Studying multiple human personality traits utilizing modern Artificial Intelligence (AI) techniques has recently gained popularity. Past studies regarding human personality assessment have used paper pencil methods, self-reports, and questionnaires. Due to the proliferation of various technologies, advancement in the Internet and use of social media networks, the concept of utilizing images, text, and videos to model human personality is gaining popularity. This work utilizes an AI-based framework to predict human personality with respect to the Big-Five model based on the choice of images made in a real-world setting. For this, the current proposal uses a dataset of real-life images that are directly/indirectly related to positive and negative sides of each of the Big-Five personality traits. Using different image seeking tasks, a data is collected from 77 participants through a custom-built tool. For creating the ground truth (about the personality of these participants), the IPIP-NEO-120 (International Personality Item Pool Representation of the NEO) personality test is taken by the participants. Results recorded are later used to correlate with the trait percentile extracted through image selection tasks. Pearson correlation coefficient is used for computing the correlation between the personality profiles of the participants. The correlation test reveals that 82% of the results are positively correlated with a p-value of 0.02. Using this data, three AI classifiers, namely, Support Vector Machine (SVM), k-nearest neighbors (k-NN), and Artificial Neural Networks (ANN) are trained for predicting the personality traits of the participants. Initially, these classifiers are trained to categorize a person being high, average, low or very low in a personality trait. Where, the maximum average accuracy of 83% is achieved by SVM for predicting agreeableness utilizing the polynomial kernel having degree six. Later, these classifiers are trained to predict the dominant personality trait, for which five class labels (i.e., O, C, E, A, and N) are assigned based on the highest percentile among all five traits. Where, SVM outperforms kNN and ANN with an average accuracy of 70%. The results reveal that different aspects of human personality can be predicted with sufficient accuracy using an individual's choice of images in a real-world setting.