Nowadays, the number of people using social network site increases every day. The social networking sites, such as Facebook or Twitter, are sources of human interaction, where users are allowed to create and share their activities, thoughts and place di erent information about themselves. However, most of this information remains unnoticed. In this work, we propose a machine learning approach to predict Big-Five personality using information from users accounts from the social network. The predictions can be used in di erent areas such as psychology, business, marketing.
Today we cannot imagine our life without social media resources. We communicate with our friends, share information with them or just spend our free time in social networks, for example on Facebook. Eventually, this process results in the accumulation of a huge amount of information, which can tell almost everything about the person, since people tend to write a lot about themselves. In our work, we will use information from Vkontakte accounts (a social network, which is popular in Russia, the analogue of Facebook) to predict information about their owners. These predictions can be used in various areas of our lives, such as psychology, business or marketing.
In the present work, we will use the psychological model of the Big Five [
Chapter 2 presents the results achieved to date in this eld. In Chapter 3 we give the main de nitions and describe predictions of the psychological scales of Big-Five and the psychological scale of Schwartz. In Chapter 4 we give the main conclusions on the work done. 2
There are a lot of articles related to the study of the Big-Five scale. One of
the earliest ones is the work written by S.D. Gosling et al. [
Later on, in [
Similar research was conducted in [
More advanced machine learning were applied in [
In our work, we use the Big-Five scales and the Schwartz scale. Big-Five scale contains 5 dimensions: { Extraversion vs. Introversion { Agreeableness vs. Antagonism { Conscientiousness vs. Lack of direction { Neuroticism vs. Sensible stability { Openness vs. Closeness to experience
Schwartz human values scale contains 10 dimensions: { Self-Direction { Stimulation { Hedonism { Achievement { Power { Security { Conformity { Tradition { Benevolence { Universalism
First of all, we consider the correlation between the initial signs and the resulting values of psychological traits. Correlation tables were constructed for the data, the most signi cant results are shown in Table 1.
city sibling grandparent parent child sex age
Further, the scale data were considered from the point of view of regression
and classi cation for constructing the predictive model. We used algorithms of
machine learning, such as Gradient boosting, Random forest, SVM, Linear /
Logistic Regression from Scikit-learn [
For regression algorithms, we compare the results presented for R2 and MSE, and accuracy, f1-score for classi cation algorithms. For each of the problems, we consider some of the best features that could be predicted (three in each case). In the case of the classi cation problem, we split our initial scale into 5 classes and try to predict the hit of an element in 1 of the classes. The results obtained are presented in Tables 2 and 3.
It can be noted that even the best results for regression are extremely poorly predicted by the signi cance of the features, which is indicated by Negative value of R2. In the case of the classi cation problem, the results are encouraging, but still the feature selection based on only social pro le is not su cient. It is also worth noting that the results obtained for the remaining characteristics do not provide su cient grounds for stating the conclusions about successful classi cation.
As a result of machine learning, the best methods were SVR, Random Forest - in the case of a regression problem, and Gradient Boosting, Logistic regression - in the case of a classi cation problem. Also, attempts were made to improve the result using SVM (the use of di erent kernels, but this transformation did not yield tangible results).
mean R2 std R2 mean MSE std MSE Self-Direction
mean f1-score std f1-score mean accuracy std accuracy Logistic Regression
Random Forest Gradient Boosting
SVM
Power Logistic Regression
Random Forest Gradient Boosting
SVM
Conformity Logistic Regression
Random Forest Gradient Boosting
SVM 0.23 0.11 0.30 0.09 0.24 0.05 0.27 0.06 0.20 0.05 0.23 0.05 0.17 0.06 0.29 0.07 mean f1-score std f1-score mean accuracy std accuracy 0.24 0.03 0.66 0.01 0.23 0.05 0.63 0.02 0.22 0.04 0.61 0.03 0.22 0.05 0.69 0.03 mean f1-score std f1-score mean accuracy std accuracy As a training dataset we used a small database of users, which were manually marked by experts. Additional veri cation was obtained via psychological tests processed by users who allowed access to their social network pro les when passing the tests. Information from accounts in the social network we will get by using VKontakte API. The information we will get contains the following characteristics: { 'status', 'about', 'wall comments', 'relation', { 'has mobile', 'has photo', 'inspired by', { 'people main', 'life main', 'political', { 'tv', 'movies', 'games', 'music', { 'smoking', 'alcohol', 'religion', { 'home town', 'city', 'languages', { 'universities','education', { 'activities', 'interests', { 'occupation', 'career', { 'followers count', { 'sex', 'age'.
We separate \complex" characteristics, such as { 'people main', 'life main', 'inspired by', { 'alcohol', 'smoking', 'political', 'religion', { 'languages', 'city', 'followers count' For the items of this group we considered the value that forms persons attitude to the given attribute (each characteristic is associated with a scale). Other attributes we associated with 0,1-values, such that 1 corresponds to the indicated attribute in the pro le, while 0 means that characteristic is not speci ed. We also added signs of sex and age in addition to numerical data. After that we form data based on the characteristics.
The distribution of result features is shown in Fig.1 First of all, a correlation table was constructed for the signs, which revealed signi cant values for some psychological signs and background data from social networks. However, for the regression problem we have received unsatisfactory results, regardless of the machine learning model. For the classi cation problem our results can be considered satisfactory. We have received predictions on certain grounds (in particular, Self-Direction, Power, Conformity). It is also worth noting that such results can be a consequence of a small sample, so we cannot unequivocally answer the question of the signi cance of the results obtained. As possible explanations for this result, there are three main reasons. Firstly, this can mean that in the future it is necessary to consider a more carefully selected set of characteristics. Secondly, it should be noted that such a poor result can be a consequence of a subjective scale of psychological signs, because tests cannot accurately determine the psychological portrait of a person. Thirdly, which is also important, the presence of unveri ed information in social accounts makes it di cult to build a working model. 6
During the work, a set of characteristics for the prediction of an account from a social network was generated, 2 models of prediction of features (regression model, classi cation model) were constructed. Predictive models were SVM, SVR, Logistic Regression, Random Forest, Gradient Boosting. As a result, significant correlations were found in some of the psychological traits with signs from social networks, and the model of prediction with classi cation was successfully reviewed. Adding information from semantics of posted text information should bene t our research in the future work.
Acknowledgements. The article was prepared within the framework of the Basic Research Program at the National Research University Higher School of Economics (HSE) and supported within the framework of a subsidy by the Russian Academic Excellence Project `5-100'.