| As a stable online content in social media,the profile image is used to show the user's persona.The faces and styles shown in the profile image will vary based on user's personality.The daily images released by users are also closely related to the user's psychological state and demographic characteristics.This article proposes the analysis and prediction of personality based on social media and the implementation of specific experiments based on the Twitter platform.Two types of the image generated by users in online activities are used to infer personality The main work focuses on user data acquisition in social media,image feature extraction and filtering,classifier filtering.Extracting and filtering image features is the focus of research.The Big Five model[11]is used to divide the personality of users.Big Five unifies individual differences in personality and divides personality into five main dimensions:agreeableness,conscientiousness,extraversion,neuroticism,and openness.First,a web crawler to is used obtain data of 1431 Twitter users who meet the experimental requirements.Then mark their personality characteristics based on the text data.Second,use two deep learning-based methods to extract facial features and styles in the profile image.This article is the first attempt to apply the image style to the personality prediction task.Third,remove redundant features,analyze the relationship between image features and personalities.After the discretization of personality,the five personality characteristics of users were predicted based on classical classification algorithms such as support vector machine,naive bayes,k-nearest neighbor and decision tree,that is,each personality characteristic corresponds to a binary classification task.In the end,when only using a picture of the user's profile image to predict personality,the prediction accuracy of neuroticism reached 77.9%,the prediction accuracy of openness is 73.7%,the prediction accuracy of agreeableness is 69.3%,the prediction accuracy of Conscientiousness is67.5%,and the prediction accuracy of extraversion is 61.4%.In addition,by analyzing the correlation,it is found that there are differences in the content of the user profile images with different personalities.For example,users with a higher degree of openness prefer to show smiley faces in profile images than users with low openness.people who lack conscientiousness are more likely to express anger in profile image.The difference in personality is also reflected in the style of the profile image.For example,people high in openness do not like‘Romantic'and‘Pastel'style.Then,the recursive feature eliminate is used to select features to enhance the performance of using the profile image to infer the personality.Results show that the recursive feature elimination method improves the performance of the personality classification model in different degrees compared with the Pearson correlation coefficient as the standard image feature selection.In order to enhance the credibility of image information and avoid the influence of single avatar on personality prediction,it is proposed to use the images published daily by users to infer personality.Due to the complexity of images published by users,it is difficult to fully consider all kinds and styles of images.In this paper,only images containing one face are selected.More importantly,the face is limited to the same person as the face in the user's profile image.A total of 156 valid users were selected manually,and on average each of them released 5 single portraits that met the limit.Results show that multiple images have significantly improved the accuracy of personality prediction.Among them,the prediction accuracy of neuroticism is 86.4%,the prediction accuracy of agreeableness is 85.7%,the prediction accuracy of openness is 83.3%,the prediction accuracy of conscientiousness is 80.0%,and the prediction accuracy of extraversion is72.1%. |
PDF Full Text Request