Home // SOTICS 2019, The Ninth International Conference on Social Media Technologies, Communication, and Informatics // View article

Identifying Latent Toxic Features on YouTube Using Non-negative Matrix Factorization

Authors:
Adewale Muyiwa Obadimu
Esther Mead
Nitin Agarwal

Keywords: Toxicity, Tonality Analysis, YouTube, Social Media, Language Model.

Abstract:
Toxic behavior, in its various forms, often disrupts constructive discussions in online communities. The proliferation of smart devices and mobile applications has further exacerbated these nefarious acts on various social media platforms. Largely, toxic behavior is regulated by human moderators employed by platform operators. However, given the volume and speed of content posted on online platforms, identifying and deterring these behaviors remains challenging. In this study, we propose a Non-negative Matrix Factorization (NMF) technique for predicting commenter toxicity on YouTube. We utilized the YouTube Data API to collect data from the Cable News Network (CNN) channel on YouTube. Our final dataset consists of 144 videos, 243,344 commenters, and 421,924 comments. We then utilized Google’s Perspective API to assign a toxicity score to each comment. We used the resultant dataset to create a commenter toxicity score prediction model. We tested our proposed NMF model against other popular prediction methods, comparing the speed of model execution and the common Root-Mean-Square-Error (RMSE) accuracy metric. This work sets the stage for a richer, more detailed analysis of toxicity on various online social media networks.

Pages: 25 to 31

Copyright: Copyright (c) IARIA, 2019

Publication date: November 24, 2019

Published in: conference

ISSN: 2326-9294

ISBN: 978-1-61208-757-3

Location: Valencia, Spain

Dates: from November 24, 2019 to November 28, 2019