Automatic identification of alcohol-related promotions on Twitter and prediction of promotion spread

摘要

Teens who have viewed alcohol-related content on social networking sites are more likely to have consumed alcohol than teens that have not seen such content. This suggests a rising concern about the influence of these sites on adolescent drinking behavior. Parents, health organizations, and school administrators need a deeper understanding of online promotional patterns in order to combat risky behaviors through intervention and education. To address these problems, we developed a system that automatically identifies alcohol promotions in online Twitter content. The identification of promotions was modeled using supervised machine learning algorithms. Predictor variables were derived from the content of tweets, the Twitter meta-data, and the network structure. We evaluated this system using held-out testing data in a cross-validated experimental design. We found that random forest models were best at predicting promotional tweets. Yet, logistic regression main effects models were useful in determining the significance of each variable, both Twitter specific and textual. For Twitter specific variables, number of hashtags and number of mentions significantly increased the likelihood of a tweet being a promotion. Using the TF-IDF method for textual predictors, we found that words that describe a type of alcohol, such as “beer” or “wine,” increased the likelihood of a tweet being a promotion. Our analysis provides information about the current state of online alcohol promotion, salient characteristics of promotions and promoters, and the influence of promotions on other users of social networking sites.