融合多维特征的电诈犯罪时空预测研究

PDF下载HTML

Title:

Spatiotemporal prediction of telecommunications network fraud crime with multidimensional feature fusion

作者:

周璟昊¹, 石磊², 石拓^3,4, 陈鹏¹

1. 中国人民公安大学 信息网络安全学院, 北京 102600;
2. 中国传媒大学 媒体融合与传播国家重点实验室, 北京 100024;
3. 北京警察学院 公安管理系, 北京 102202;
4. 北京警察学院 北京市公安局警察学院警务情报与数据智能标准实验室, 北京 102202

Author(s):

ZHOU Jinghao¹, SHI Lei², SHI Tuo^3,4, CHEN Peng¹

1. School of Information Network Security, People’s Public Security University of China, Beijing 102600, China;
2. State Key Laboratory of Media Integration and Communication, Communication University of China, Beijing 100024, China;
3. Department of Public Security Management, Beijing Police College, Beijing 102202, China;
4. Standard Laboratory of Police Data and Intelligence of Beijing Public Security Bureau, Beijing Police College, Beijing 102202, China

关键词:

电诈犯罪时空预测; 多维特征; 时空特征融合; 空间环境特征因子; 图注意力网络; 时间滑动窗口; iTransformer

Keywords:

spatiotemporal fraud prediction; multidimensional features; spatio-temporal feature fusion; spatial environmental feature factors; graph attention networks; time-sliding window; iTransformer

分类号:

TP391.41

DOI:

10.11992/tis.202412025

摘要:

对空间内电信网络诈骗犯罪发案进行预测能够有效提升精准反诈工作效能。然而，现有方法受到发案时序数据存在稀疏性和周期性的影响，以及空间环境特征因子异质性限制的挑战，导致面向电信网络诈骗发案量的预测效果不佳。为此，提出一种融合多维特征的电信网络诈骗犯罪时空预测模型(multidimensional feature-integrated telecom fraud spatiotemporal prediction model, MF-TSP)。结合区域空间拓扑图构建空间环境特征因子筛选模块，有效融合预测目标空间的邻域发案特征；运用时间滑动窗口技术，并引入多维时序特征捕捉模块和倒置Transformer(inverted transformers, iTransformer)模块，克服了发案时序数据稀疏问题，同步实现了对序列周期性、全局依赖关系及多变量间复杂相关性的有效捕捉；通过进一步深度时空特征融合和非线性映射，显著提升了犯罪发案量的预测精度。实验结果表明，提出的MF-TSP模型在B市电信网络诈骗犯罪发案真实数据集上，在3种不同输入时间步长条件下均表现最佳，明显优于7种对比模型。

Abstract:

Spatiotemporal prediction of telecommunications fraud crimes can substantially enhance targeted antifraud efforts. However, existing methods suffer from poor performance due to sparse and periodic incident time-series data, as well as the heterogeneity of spatial environmental factors. Aiming to address these challenges, this paper proposes a multidimensional feature-integrated telecom fraud spatiotemporal prediction (MF-TSP) model. First, a spatial feature selection module was constructed by integrating regional topological graphs to effectively incorporate neighborhood crime patterns. A time-sliding window technique, combined with a multidimensional temporal feature extraction module and an inverted Transformer, addresses data sparsity while capturing periodicity, global dependencies, and complex multivariate correlations. Furthermore, deep spatiotemporal fusion and nonlinear mapping notably improve prediction accuracy. Experiments on real-world telecom fraud data from City B demonstrate that MF-TSP outperforms seven baseline models under three different input time-step conditions.