Abstract: At present, terrorism has become an important factor affecting world peace and development. Analyzing terrorist attacks to extract useful information and predict the spatial behavior of terrorist organizations has become one of the hotspots of current research. Although there have been many studies on the spatial prediction of terrorist attacks, most methods still have room for improvement in constructing the spatial behavior relationship network of terrorist organizations and extracting higher-order relationships. In order to solve this problem, the spatial behavior prediction problem of terrorist organization is modeled as a collaborative knowledge graph composed of terrorist organization-attack site interaction network and attack site knowledge graph. This paper proposed a spatial behavior prediction method KGCE(Knowledge Graph Attention Network for Recommendation) of terrorist organization based on improved KGAT TransE). Firstly, according to the interactive relationship between the terrorist organization and the attack site and the knowledge graph of the attack site, the terrorist organization-attack site interaction network and the attack site knowledge graph are constructed respectively. By combining the two networks, a terrorist organization spatial behavior collaborative knowledge graph based on socio-spatial relationship is proposed. Secondly, the structure of KGAT is improved, and the TransE model is introduced into the embedding layer to alleviate the problem of model overfitting. At the same time, the overall model effectively models the high-order relationship in the collaborative knowledge graph of spatial behavior of terrorist organizations in an end-to-end manner. By comparing with five mainstream competitive baselines on the public data set, the experimental results and analysis show that KGCE is higher than the existing baselines in the prediction accuracy of terrorist organization spatial behavior, and its recall rate is improved by up to 7.14%, which verifies the effectiveness and correctness of the proposed framework.

Key words: Terrorist organizations, Attention mechanism, Spatial behavior prediction, Recommendation system

摘要： 当前，恐怖主义已成为影响世界和平与发展的重要因素。分析恐怖袭击事件以提取有用信息，并预测恐怖组织的空间行为成为当下研究的热点之一。虽然已经有较多对于恐怖袭击空间预测的研究，但大多数方法在构建恐怖组织空间行为关系网络并显示提取高阶关系方面仍有提升空间。为解决这一问题，对恐怖组织空间行为预测问题建模,为由恐怖组织-袭击地点交互网络和袭击地点知识图构成的恐怖组织空间行为协同知识图，提出了一种基于改进KGAT（Knowledge Graph Attention Network for Recommendation）的恐怖组织空间行为预测方法KGCE(Knowledge Graph Recommendation of TransE)。该方法首先根据恐怖组织与袭击地点之间的交互关系及袭击地点的知识图谱，分别构建了恐怖组织-袭击地点交互网络及袭击地点知识图，通过将这两种网络结合，提出了一种基于社会-空间关系的恐怖组织空间行为协同知识图；其次改进了KGAT结构，在嵌入层中引入TransE模型以缓解模型过拟合问题，同时整体模型以端到端的方式，有效地对恐怖组织空间行为协同知识图中的高阶关系进行了建模。在公开数据集上通过与五个主流竞争基线进行对比，实验结果和分析表明，KGCE在恐怖组织空间行为预测准确度方面高于现有基线，其召回率最高提升7.14%，验证了本文框架的有效性与正确性。

关键词: 恐怖组织, 注意力机制, 空间行为预测, 推荐系统