基于Lightgbm算法的恐怖袭击预警模型构建研究

Research on the Construction of Terrorist Attack Early Warning Model Based on Lightgbm Algorithm

[研究目的]在GTD海量恐怖主义活动数据中,存在特征影响恐怖袭击目标实现,使用机器学习方法建立恐怖袭击预警模型,可为反恐预警提供决策支持。[研究方法]通过预测恐怖袭击目标实现风险,可发现恐袭预警重要性特征。对135维GTD特征进行特征筛选、归一化、独热编码、卡方检验、PCA降维,利用Lightgbm算法在内四种机器学习算法测试评估,并根据Lightgbm算法特征重要性,控制变量并重复实验得到重点及突增点特征。[研究结果]综合评价指标,证明Lightgbm算法在表现上优于其他机器学习算法,在保证样本类别1∶1均衡的前提下,910次实验其平均准确率为0.7986,召回率为0.7852,f1值为0.7832,平均运行时间为3.57s。Lighgbm算法在GTD数据上可有效提高分类效果,attacktype突增点特征及前十四顺位特征应作为重点特征辅助预警决策。

[Research purpose]In GTD's massive terrorist activity data,there are features affect the realization of terrorist attack targets.Using machine learning method to establish early warning model of terrorist attack can provide decision-making support for anti-terrorism early warning.[Research method]This study predicts the realization risk of terrorist attack targets.So the importance of early warning of terrorist attacks can be found.It performs feature filtering,normalization,one-hot encoding,chi-square test,PCA dimensionality reduction on 135-dimensional GTD features,and uses four machine learning algorithms including the Lightgbm algorithm to evaluate the results.According to the feature importance of lightgbm algorithm,we control the variables and repeat the experiment to get the key features and sudden increase feature.[Research conclusion]The comprehensive evaluation index proves that lightgbm algorithm is superior to other machine learning algorithms in performance.Under the premise of ensuring 1:1 balance of sample categories,the average accuracy of 910 experiments is 0.7986,the recall rate is 0.7852,the f1 value is 0.7832,and the average running time is 3.57s.The lighgbm algorithm can effectively improve the classification effect on GTD data.The sudden increase point feature“attacktype”and the top fourteen ranking features should be used as key features to assist early warning decision-making.