A feasibility study on identifying drinking-related contents in Facebook through mining heterogeneous data

Area of Science:

• Social Sciences
• Computer Science
• Public Health

Background:

• Binge drinking is a significant public health issue among US college students.
• Social media platforms are frequently used by students to share alcohol-related content, often normalizing excessive consumption.

Purpose of the Study:

• To assess the feasibility of using machine learning to mine heterogeneous social media data (text, images, videos) for identifying college student binge drinking.
• To evaluate the effectiveness of different data types in predicting drinking-related content.

Main Methods:

• Manual annotation of 4266 Facebook posts from the "I'm Shmacked" group between October 2011 and November 2014.
• Development and evaluation of machine learning models utilizing combined and individual data types (text, image, video) to classify drinking-related posts.

Main Results:

• Out of 4266 annotated posts, 511 were identified as drinking-related.
• Machine learning models combining heterogeneous data achieved an F1-score of 0.81 in identifying drinking-related posts.
• Text-based prediction models demonstrated higher reliability compared to image and video data.

Conclusions:

• Mining social media data presents a promising approach for identifying college students engaging in binge drinking.
• This feasibility study highlights the potential of leveraging multimodal social media analysis for public health interventions related to alcohol consumption.