Introduction
The COVID-19 pandemic necessitated a rapid global transition to digital platforms for social, educational, and professional interactions. While these technologies maintained societal continuity, they also expanded the environment for online incivility and harassment. This study investigates whether the pandemic triggered a measurable surge in cyberbullying by analyzing nearly half a million tweets from early 2020. Using change-point detection algorithms, the research identifies statistically significant shifts in the volume of cyberbullying-related discussions and correlates these shifts with the timeline of the global health crisis.
Methodology
The researchers employed a combination of data mining and time-series statistical analysis:
- Data Acquisition: Historical data from the first half of 2020 was collected using the "Get Old Tweets" API to circumvent standard Twitter API limitations.
- Categorization: Keywords were grouped into three distinct sub-classes: CY (terms containing "cyber"), ON (terms containing "online/internet"), and TW (terms containing "twitter").
- Change-Point Analysis: The study utilized the At Most One Change (AMOC) model, specifically applying a Cumulative Sum (CUSUM) technique. This algorithm identifies the specific point ((\tau)) where the mean of a time series shifts significantly. Analysis was conducted using the
changepointpackage in R, with a 95% confidence level ((\alpha = 0.05)).
Dataset Overview
The study analyzed a custom dataset of 454,046 unique public tweets posted between January 1st and June 7th, 2020. The search criteria included 28 specific keywords such as cyberbullying, online harassment, and internet bullies. To ensure the daily count reflected unique discourse rather than viral repetition, retweets and duplicates were filtered out.
Analysis and Key Insights
The statistical analysis revealed several critical shifts in online behavior:
- The Lockdown Correlation: For the "Cyber" (CY) sub-class—representing the most direct terminology for the phenomenon—the statistical changepoint occurred on March 29th. This date aligns closely with the implementation of strict stay-at-home orders in many countries.
- Sub-class Divergence: Different keyword categories reacted at different speeds. While CY terms shifted in late March, Twitter-specific bullying terms (TW) did not see a significant mean shift until April 27th, suggesting that platform-specific terminology may lag behind general crisis discourse.
- The "Hana Kimura" Effect: A massive, non-COVID-related spike was observed in late May. This was attributed to the tragic suicide of a Japanese reality TV star following intense online abuse. This indicates that while the pandemic raised the baseline of harassment, specific high-profile events still drive the most significant conversational peaks.
Results
The analysis substantiated the hypothesis that the pandemic increased cyberbullying discussions. Key results include:
- CY Sub-class: Changepoint identified on March 29th (235,542 tweets).
- TW Sub-class: Changepoint identified on April 27th (96,147 tweets).
- Specific Keywords: Terms like "cyber bully" and "cyberbullies" saw their statistical break on March 31st.
Limitations and Future Directions
The authors noted that the current AMOC model is limited, as a long-term crisis likely involves multiple change points. Furthermore, autocorrelation tests (ACF plots) revealed that tweet counts are not independent, which may affect the precision of changepoint locations. Future research will focus on Bayesian time-varying models and count autoregressive series to improve accuracy and potentially predict future spikes in online harassment, enabling platforms to implement proactive mitigation strategies.
Summary updated