Scholarlink

<p>This study investigates whether machine learning models trained on social media data can reliably predict real-world Black Lives Matter (BLM) protests in New York City, with a focus on the importance of temporal dynamics in the online-offline relationship. Using the Elephrame protest database and Giorgi's Twitter Corpus covering August 2014 to December 2020 (239 protests; 28,061 tweets), we apply temporal data splitting to prevent data leakage and compare feature engineering strategies across seven algorithms while addressing class imbalance. The findings reveal that simple models with well-chosen features can reliably predict protest events. Notably, hashtag volume dynamics are significantly more predictive than semantic content, indicating that engagement metrics capture mobilization signals more effectively than text analysis. The results also demonstrate concept drift, showing that the digital signal patterns associated with protests evolve over time. Overall, the results show that social media discourse can effectively predict offline protests. They support collective action theory’s emphasis on mobilization and political process theory’s focus on planning capacity. However, concept drift suggests that future predictive systems must be dynamic and adaptive to remain effective.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Concept drift in event prediction from social media

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Naicheng Wei,
  • Shlomo Engelson Argamon

摘要

This study investigates whether machine learning models trained on social media data can reliably predict real-world Black Lives Matter (BLM) protests in New York City, with a focus on the importance of temporal dynamics in the online-offline relationship. Using the Elephrame protest database and Giorgi's Twitter Corpus covering August 2014 to December 2020 (239 protests; 28,061 tweets), we apply temporal data splitting to prevent data leakage and compare feature engineering strategies across seven algorithms while addressing class imbalance. The findings reveal that simple models with well-chosen features can reliably predict protest events. Notably, hashtag volume dynamics are significantly more predictive than semantic content, indicating that engagement metrics capture mobilization signals more effectively than text analysis. The results also demonstrate concept drift, showing that the digital signal patterns associated with protests evolve over time. Overall, the results show that social media discourse can effectively predict offline protests. They support collective action theory’s emphasis on mobilization and political process theory’s focus on planning capacity. However, concept drift suggests that future predictive systems must be dynamic and adaptive to remain effective.