Crime Prediction Technology Based On Multi-view And Multi-source Data Fusion

Criminal activities threaten the personal and property security of citizens and the harmony and stability of society.How to prevent crime has become one of the challenges of social governance.Crime prediction technology can mine potential crime patterns from historical crime data and predict future crime situations.The prediction results can help to develop anti-crime policies and optimize the patrol routes of the police.There are two current challenges in the field of crime prediction.The first challenge:the features of the three most popular perspectives(temporal,spatial and type)are extracted independently in turn,and the other two perspectives are ignored when one of them is processed,so that the feature extraction and fusion in this way may lead to the loss of useful information.The second challenge:some scholars have introduced some auxiliary data in addition to crime data,such as point-of-interest data,urban anomaly data and cab traffic data.These data can laterally reflect the internal situation of jurisdictions and the spatial association of different jurisdictions,providing an explanatory basis for the causes and spread of criminal activities.However,the current research work is too simple in the way of fusing these data,and directly overlaying or stitching them together may not bring out a better fusion effect.To address the first challenge,this thesis proposes a multi-perspective feature extraction and fusion technique.Breaking the feature extraction order existing between the three perspectives of time,space and type,a multi-view fusion graph is built based on type nodes and bridged by time and space.On top of this graph,this thesis proposes a multi-view simultaneous convolution algorithm,where the convolution process is able to extract and fuse the features of the three perspectives of time,space and type simultaneously.To address the second challenge,this thesis proposes a multi-source data fusion technique.The data reflecting the spatial connections between different jurisdictions are fused into the multi-view fusion map.The data reflecting the internal situation of jurisdictions are processed and mapped to the data space of crime data,so that several additional introduced auxiliary data and crime data have nearly the same data distribution to achieve better fusion effect.The two techniques proposed in this thesis are fully experimented on several realworld crime datasets as well as POI datasets,urban anomaly datasets and cab traffic datasets.The first research content(multi-view feature extraction and fusion technique)improves up to 22%and 5.2%on the Macro-F1 metric compared to the two newer crime prediction models,Mist and CrimeForecaster.The second research content(multisource data fusion technique)improves 3.1%on average over the first research content.In addition to this,we also conducted ablation experiments on both techniques to fully validate the necessity of each component module.