摘要
随着大数据时代的来临,各领域的研究越来越依托于大数据,机器学习算法已成为社会科学研究的热门工具之一。在和平与冲突研究领域,作为预测冲突的重要手段,机器学习算法认为通过大量数据训练得到的模型能够准确预测国家的冲突行为和冲突事件。一些项目并没有事前公开其数据、算法和代码,因此无法对其预测能力进行评估。另一些项目和研究虽然公布了它们的数据和算法等,但在关键指标上表现欠佳。通过检验暴力预警系统,发现该系统基于机器学习的相关分析对非洲地区冲突的预测效果并不理想。事实上,无论是基于统计学建立的因果关系模型还是机器学习挖掘出的关联性规则,都无法精准预测未来发生的冲突,其原因在于国家的内外部环境会随着时间推移而发生变化。新冲突的发生可能受全新因素的影响,也可能受因素之间相互作用的影响,还可能是由不同机制驱动所致,但这些因素都无法被机器学习与算法所捕捉,这导致相关模型和算法缺乏外部效度。通过机器学习算法的方式预测冲突这一涌现性结果的前景并不乐观,研究者在对冲突进行预测时还需要融合多种方法。
With the advent of the big data era,research in various fields has increasingly relied on big data,and machine learning algorithms have become a popular tool in social science research.In the field of peace and conflict studies,machine learning has become an important research method for predicting conflicts.It is believed that models trained on large amounts of data can accurately predict a country's conflict behavior and conflict events.In projects and studies that support policy-making,some do not publicly disclose their data,algorithms and code,making it impossible to evaluate their predictive capabilities.Other projects or studies that have disclosed their data and algorithms may not perform well on key indicators.By examining the Violence Early-Warning System(ViEWS),it was found that the system's machine learning-based analysis had unsatisfactory performance in predicting conflicts in Africa.In fact,neither causal models established based on statistical models(conflict factors and conflict eruption)nor association rules discovered through machine learning can accurately predict future conflicts.This is because the internal and external environment of a country can change over time.The occurrence of new conflicts may be influenced by entirely new factors or the interaction between factors,or driven by different mechanisms,which cannot be captured by machine learning algorithms,leading to a lack of external validity in the related models and algorithms.The prospects of using machine learning algorithms to predict conflict outbreaks,an emergent result,are not optimistic.The reasons behind this are not only methodological issues but also deeper ontological and epistemological problems.Therefore,a combination of multiple methods is needed when predicting future conflicts.
作者
刘辰辉
唐世平
Liu Chenhui;Tang Shiping
出处
《世界经济与政治》
北大核心
2023年第12期114-143,171,172,共32页
World Economics and Politics
基金
2019年上海市哲学社会科学规划青年课题“非洲国家族群冲突风险预测”(项目批准号:2019EGJ002)的阶段性成果。
关键词
机器学习
冲突预测
社会结果
非洲地区冲突
行动
machine learning
conflict
prediction
social result
regional conflicts in Africa
