Analysis Method Of Drug-Related Personnel Association Relationship Based On Graph Convolutional Neural Networks

In recent years,the rapid evolution of drug networks and the intricate relationships among drug-involved individuals have posed challenges to public security agencies’ efforts in combating drug-related crimes.Particularly,with the advancement of communication technologies,the clandestine nature of drug networks has intensified,rendering traditional prevention and control strategies increasingly inadequate.Therefore,the construction of a scientifically grounded network model for drug-involved personnel association relationships,utilizing deep learning techniques to thoroughly analyze and predict drug-related behaviors and their relationships,has emerged as an urgent necessity for effectively intervening and severing the drug crime chains.This dissertation has been grounded in the needs of public security agencies in combating drug-related crimes,focusing on the analysis of association relationships among drug-involved personnel.It integrates multidisciplinary theories and technologies such as social network analysis,graph convolutional neural networks,and spatiotemporal data mining.Taking complex drug networks and drug-involved individuals as research objects,this dissertation systematically investigates the framework and methods for drug-involved personnel and behaviors identification,association relationship analysis,and prediction of drug-related behaviors.Through the construction of a system of characterization indicators for drug-involved individuals and a network model of association relationships among drug-involved individuals,it aims to provide public security agencies with more precise tools for analyzing drug cases,thereby bearing significant theoretical significance and practical value.The innovations and contributions of this dissertation are as follows:(1)This dissertation has established a set of characterization indicators for drug-involved individuals,conducting in-depth research to identify individual and behavioral patterns within the drug-involved personnel network,effectively describing their features.The indicator system has comprehensively considered both the ontological and behavioral features of drug-involved personnel,encompassing 55 dimensions of features including personal background,social network relationships,criminal history,and behavioral patterns.Drawing on interdisciplinary theories such as sociology,information dissemination,homogeneity,power-law distribution,and spatiotemporal network analysis,this study has conducted a thorough analysis of the complex correlations within drug networks.Utilizing Pearson correlation coefficients and expert rating methods to quantify and grade the features of drug-involved personnel has enhanced identification and classification accuracy,and uncovered 9 potential association relationships among drug-involved individuals,revealing complex interaction patterns and influencing mechanisms.Additionally,a dataset of drug-involved personnel networks(DRN)has been created.The aforementioned research findings provide a theoretical foundation for constructing potential drug-involved personnel networks and establish a data foundation for subsequent research endeavors.(2)This dissertation has proposed a method for constructing a potential drug-involved personnel network based on deep graph convolutional neural networks,addressing the issues of determining the scope of potential drug-involved personnel,formulating completion standards and methods,and predicting whether potential drug-involved individuals have association links in the completed drug network.This method has complemented the captured drug-involved personnel network based on mined potential drug-related relationships and has introduced the Υdecay theory to determine the completion scope of potential drug-involved personnel.A novel end-to-end deep graph convolutional neural network model and a non-local message passing framework have been proposed to learn the representation of drug-involved personnel subnetwork features.Furthermore,a framework for predicting potential drug-involved personnel association links based on second-order subgraph analysis has been introduced to determine the existence of association links among completed potential drug-involved personnel.Through these methods,a comprehensive and accurate potential drug-involved personnel network has been constructed.Experimental results have demonstrated that the proposed method significantly improves the accuracy of graph classification tasks by approximately 24.26%-43.44% compared to numerous baseline methods on four datasets and has achieved a prediction accuracy of over 98% in association link prediction on drug-involved personnel network datasets.(3)This dissertation has proposed a method for identifying and classifying drug-involved individuals based on hyperbolic deep graph convolutional neural networks,addressing the inadequacies of traditional methods in handling the power-law distribution characteristics and complex hierarchical structures of drug networks.This method has transferred the druginvolved personnel network from Euclidean space to hyperbolic space,leveraging the hierarchical and geometric properties of hyperbolic space for more accurate learning of hierarchical features within the drug-involved personnel network.Core operations of deep graph convolutional neural networks have been defined in hyperbolic space and combined with innovative hyperbolic feature transformation methods based on identity mapping and non-local message passing frameworks,enabling effective exploration and analysis of remote dependency relationships and drug-involved personnel features within the network.An innovative neighborhood aggregation method on hyperbolic manifolds has been proposed,further enhancing the model’s performance by integrating the initial structural features of the druginvolved personnel network with attention mechanisms in hyperbolic space.Experimental results have demonstrated that the proposed method achieves an average increase of approximately 8.13% in the F1 score for drug-involved personnel classification on druginvolved personnel network datasets compared to other baseline hyperbolic graph neural network methods.(4)This dissertation presents a method for analyzing drug-involved personnel association relationships based on multi-dimensional edge-embedding deep graph convolutional neural networks,addressing the challenge of effectively learning the features of drug-involved personnel association relationships with traditional social network analysis methods.This method utilizes deep learning techniques to uncover the deep features of association relationships,thereby effectively revealing hidden connections among drug-involved individuals.It introduces a novel message passing framework capable of transmitting multidimensional edge information,facilitating the comprehensive integration and utilization of drug-involved personnel features and multi-dimensional association relationship features.Additionally,a new graph convolutional layer has been proposed,enabling simultaneous learning of drug-involved personnel feature embeddings and multi-dimensional association relationship feature embeddings in each layer,thus achieving efficient utilization and fusion of both features.Furthermore,a method for encoding multi-dimensional edge features based on potential association relationships among drug-involved personnel has been proposed,effectively improving the accuracy of drug-involved personnel association relationship identification and classification.Experimental results on drug-involved personnel network datasets demonstrate that the proposed method outperforms state-of-the-art methods,achieving improvements of 4.47%,5.88%,and 1.77% in association relationship classification accuracy,respectively.(5)This dissertation proposes a method for predicting drug-involved behaviors of drug users based on deep spatiotemporal graph convolutional neural networks,addressing the integration inadequacies of traditional predictive models in dealing with complex drug networks’ temporal and spatial analyses,as well as the challenge of effectively modeling dynamic spatiotemporal correlations in temporal data.This method introduces a deep spatiotemporal graph convolutional neural network based on the temporal-spatial dynamic features of drug-involved personnel networks and embeddings of multidimensional latent association relationship features,enabling the analysis and prediction of drug-involved behaviors’ temporal-spatial dynamics.It also proposes a method for multiscale periodic analysis of drug-involved personnel networks,extracting three types of time characteristics,including yearly,monthly,and weekly cycles,to fully utilize temporal dynamic information.Additionally,it designs a spatiotemporal graph convolutional layer based on spatiotemporal attention mechanisms for drug-involved personnel and association relationship features,where spatial graph convolution effectively captures non-local spatial features of drug-involved personnel,and temporal convolution adequately extracts dynamic temporal correlation features.Experimental results on drug-involved personnel network datasets demonstrate that compared to six baseline methods,the proposed method achieves an average decrease of 2.87 and 4.01 in MAE and RMSE indicators,respectively.