CONcISE

Introduction

CONcISE is a novel two-stage online approach for timely, scalable and accurate cyberbullying detection on Instagram.

Citation

To cite your paper, please use the following reference:

Mengfan Yao, Charalampos Chelmis, and Daphney Zois. Cyberbullying Ends Here: Towards Robust Detection of Cyberbullying in Social Media. The 2019 World Wide Web Conference (WWW ’19), May 2019.

BibTeX:

@inproceedings{yao2019cyberbullying,
  title = {Cyberbullying Ends Here: Towards Robust Detection of Cyberbullying in Social Media},
  author = {Yao, Mengfan and Chelmis, Charalampos and Zois, Daphney-Stavroula},
  booktitle = {The World Wide Web Conference},
  series = {WWW '19},
  pages = {3427--3433},
  year = {2019},
  isbn = {978-1-4503-6674-8},
  url = {http://doi.acm.org/10.1145/3308558.3313462},
  doi = {10.1145/3308558.3313462},
  organization={ACM}
}

Dataset

We use a corpus of 10 thousand comments, manually annotated as harassing or not by 10 experts. We focus on harassing comments due to their commonality to a number of types of unwanted behavior, including cyberharassment and cyberbullying. We used this data to evaluate our method for robust and timely detection of cyberbullying as well as for harashment anticipation.

Our labeled corpus spans 22.1% of all media sessions containing at least 40% profanities in the dataset introduced by this paper. Of all media sessions containing at least 40% profanities, 47.5% had been labeled as positive if:

there are negative words and/or comments with intent to harm someone or other, and the posts include two or more repeated negativity against a victim.

Dataset Access

Email us at cchelmis@albany.edu if you are interested in our dataset! We are happy to share our data with you.

Prerequisites

python 2.7 and the following libraries

pandas
numpy
os
statsmodels

Files

CONcISE.py: main function of the approach
train.csv: a sample training data 
test.csv: a sample testing data