The Bag of Communities: Identifying Abusive Behavior Online with Preexisting Internet Data
summary
Introduces the Bag of Communities (BoC) method that detects abusive posts by measuring a post's similarity to large-scale data from other online communities (4chan, Reddit, Voat, MetaFilter). Shows BoC can classify abuse in a target community without any in-domain training (~75% accuracy) and that a dynamic BoC trained with 100,000 moderated posts reaches 91.18% accuracy, outperforming traditional in-domain methods.
AI-written summary · about
chapters (8)
- 0:00Start
- 2:25Examples of Abusive Comments
- 7:09Source Communities: Hypothesized Abuse
- 12:33Implications
- 12:53No need to start from scratch to detect abuse
- 13:10Boost abuse detection on established sites
- 13:20Future Work Extend to multiple sources and targets
- 13:53The Bag of Communities Identifying Abusive Behavior Online with Preexisting Internet Data
related talks
29mSocial Media Botnet Detection with Linkage Analysis and Machine Learning
37mData Driven Diagnosis of Alzheimer's Disease
38mEthical Artificial Intelligence to Advance Welcoming Communities and Identify Potential Communica...
37mData Driven Diagnosis of Alzheimer's Disease
57mTackling toxicity: How to build a thriving player community
44mApplied Machine Learning for Data Exfil and Other Fun Topics
23mBot or human? Detecting malicious bots with machine...
50mDeep Impact: Recognizing Unknown Malicious Activities from Zero Knowledge