TGBully - Temporal Graph-based Cyberbullying Detection

Implementation of Temporal Graph-based Cyberbullying Detection Model [1].

The TGBully framework consists of three major components: (1) a semantic context modeling module that encodes each comment by considering both its textual content and user’s language behavior reflected from her/his historical comments, (2) a temporal graph interaction learning module that constructs a temporal graph and models the dynamic user interaction with a bully-featured GAT. The proposed GAT jointly captures topic coherence and temporal dynamics in the modeling process; and (3) a session classification module that attentively aggregates information from user interaction into a session representation, based on which it then classifies the session into a bullying/non-bullying session.

Code Usage

1. Download all the required data and word embedding files as instructed by the data README.

2. Run the script Instagram_code.ipynb and Vine_code.ipynb for experimental results on the two benchmark datasets.

Environment

• python == 3.6.7
• cudnn == 7.1.2
• keras == 2.2.4
• numpy == 1.16.4
• tensorflow-gpu == 1.12.0

Reference

[1] Suyu Ge, Lu Cheng, and Huan Liu. Improving Cyberbullying Detection with User Interaction. The Web Conference (WWW), 2021. [paper]