A Study of Arabic Social Media Users—Posting Behavior and Author’s Gender Prediction

Social media opens up numerous possibilities to study human interaction and collective behavior in an unprecedented scale. It opened a whole new venue for research under the name “social computing”. Researchers are interested in profiling individuals (e.g., gender, age group), groups, community, and networking. We are interested in studying the collective behavior of Arabic social media users. Most studies covering Arabic social media has focused on the sentiment analysis of, say tweets. This study, however, looks into who and when users interact with the Arabic social media. Specifically, there are two objectives of this work. First, studying the demographic posting behavior of social media users from two different perspectives: gender and educational level. Second, author profiling. Identifying author’s gender of a social media post. We use Saudi Arabia, a very prolific country when it comes to social media in general, as a backdrop for this study. The results in this study are based on mining huge amount of metadata of a popular local social media forum covering the period 2011–14 inclusive. The extracted features (normalized list of k highest scoring words, and likewise for stems) from the posts were used to train classifiers to identify the author’s gender. We used two different classifiers, Support Vector Machine (SVM) with linear kernel and 1-NN (1-nearest neighbor), and experimented with different sizes for the list of features. When the number of features (size of the features vector) is small (≤ 50) both classifiers perform equally well in identifying the author’s gender, but we risk overfitting the data. The classifiers achieved their best result when using 100 features. The 1-NN classifier delivered a better performance, achieving a balanced accuracy of 93.16% vs 87.33% for the SVM in predicting the author’s gender. And for a larger set of features, SVM delivered a better performance and more stable behavior than 1-NN, but still nowhere close to its best performance. We used t test to confirm our assessment that the difference between the performance of both classifiers is statistically significant. Based on that, we recommend using 100 features, and we get our best result using 1-NN with a balanced accuracy of 93.16%.