多尺度视角特征动态融合的盗窃犯罪预测模型

PDF下载HTML

Title:

Prediction model of theft crime based on the dynamic fusion of multiscale perspective characteristics

作者:

石拓^1,2, 张齐^1,2, 石磊³

1. 北京警察学院 公安管理系，北京 102202;
2. 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;
3. 中国传媒大学 媒体融合与传播国家重点实验室，北京 100024

Author(s):

SHI Tuo^1,2, ZHANG Qi^1,2, SHI Lei³

1. Department of Public Security Management, Beijing Police College, Beijing 102202, China;
2. Standard Laboratory of Police Data and Intelligence of Beijing Public Security Bureau, Beijing Police College, Beijing 102202, China;
3. State Key Laboratory of Media Integration and Communication, Communication University of China, Beijing 100024, China

关键词:

犯罪预测; 自注意力机制; 多尺度特征融合; 卷积神经网络; 动态自适应; 分类器; 时序预测; 分布式表征

Keywords:

crime prediction; self-attention mechanism; multiscale feature fusion; convolutional neural networks; dynamic adaptation; classifier; time series prediction; distributed representation

分类号:

TP391

DOI:

10.11992/tis.202203016

文献标志码:

2022-08-23

摘要:

针对盗窃犯罪时空预测特征融合不精、时序动态适应性不足问题，提出自注意力和多尺度多视角特征动态融合的预测模型。首先，以盗窃发案的位置信息为基础，将数据投射到地图栅格内，通过构建一种可将不同时序长度案件数据匹配为自适应长度数据的方法，并组合向量映射后的天气、作案时间、地理位置等属性，构造多维度特征融合的输入向量；其次，采用自注意力机制生成多视角特征动态融合的向量；最后，通过采用多尺度窗口CNN对多视角特征动态融合向量进行编码后送入分类器，预测出每个地图栅格内的发案态势。在某市盗窃数据集上验证，本文方法在3种地理栅格尺度下，预测准确率最高可达到0.899，显著优于其他对比模型。

Abstract:

A prediction model combining self-attention and dynamic fusion of multiscale and multiview features is proposed to solve the problems of inaccurate fusion of spatiotemporal prediction features and insufficient temporal dynamic adaptability of theft crime. Initially, data are processed by constructing a method that can match case data with different lengths of time series to an adaptive length by projecting the crime data onto the map grid based on local longitude and latitude information. After word vector mapping, the weather, crime time, and location are used to construct the input vector of multidimensional feature fusion. In addition, a self-attention mechanism is introduced to generate the vector of a dynamic fusion of multiview features. The final step involves encoding the dynamic fusion vector of perspective features and sending it to the classifier to predict the crime situation in each map grid. By validating the method on a real dataset of theft crimes in a city, the proposed model can achieve a maximum prediction precision of 0.899 at three different geographic grid divisions, which is significantly better than other comparable models.