Scholarlink

<p>Political violence, including targeted attacks on women, presents significant threats to global stability and human security, yet remains an underexplored domain in time-series forecasting research. This research studies the applicability of state-of-the-art time-series forecasting methods for predicting Political Violence Targeting Women (PVTW) events and fatalities. Leveraging deep learning advancements, we evaluate state-of-the-art methods, including transformer-based architectures, traditional multi-layer perceptron models, and linear approaches such as DLinear, in the context of PVTW data. The analysis highlights the unique temporal patterns of low-intensity and high-intensity events, demonstrating that while transformer-based models outperform linear models overall, simpler architectures, such as the vanilla transformer, often match or exceed the performance of more advanced models like AutoFormer. Building on these insights, we propose a magnitude-decomposition Transformer designed to incorporate domain-specific characteristics of conflict dynamics. The proposed model effectively captures both the smaller, regular patterns and the rarer, high-intensity events in PVTW data. The findings underscore the importance of adapting deep learning architectures to domain-specific challenges and provide critical insights for designing targeted interventions and policies to address societal challenges.</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

Time-series forecasting for political violence targeting women

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月9日

查看全文
  • Myo Thida

摘要

Political violence, including targeted attacks on women, presents significant threats to global stability and human security, yet remains an underexplored domain in time-series forecasting research. This research studies the applicability of state-of-the-art time-series forecasting methods for predicting Political Violence Targeting Women (PVTW) events and fatalities. Leveraging deep learning advancements, we evaluate state-of-the-art methods, including transformer-based architectures, traditional multi-layer perceptron models, and linear approaches such as DLinear, in the context of PVTW data. The analysis highlights the unique temporal patterns of low-intensity and high-intensity events, demonstrating that while transformer-based models outperform linear models overall, simpler architectures, such as the vanilla transformer, often match or exceed the performance of more advanced models like AutoFormer. Building on these insights, we propose a magnitude-decomposition Transformer designed to incorporate domain-specific characteristics of conflict dynamics. The proposed model effectively captures both the smaller, regular patterns and the rarer, high-intensity events in PVTW data. The findings underscore the importance of adapting deep learning architectures to domain-specific challenges and provide critical insights for designing targeted interventions and policies to address societal challenges.