Research On Intelligent Risk Prediction Methods For Digital Financial Fraud

Financial fraud is a kind of illegal means to obtain economic benefits.The risk of financial fraud seriously damages the operating efficiency and stability of the capital market,and has a severe impact on the economic market.The new generation of information technology enables traditional financial industry to gradually transform into digital finance.However,the means of financial fraud continue to escalate.Frequently traded financial data is high-dimensional,complex,and nonlinear,which makes digital financial fraud risk prediction face new challenges.Traditional anti-fraud methods have a single dimension.Their computational efficiency is low,making it difficult to effectively serve the increasingly sinking user groups.Hence,the financial market urgently needs to employ intelligent risk prediction technology to improve the ability to identify digital financial fraud risks,and to predict fraud risks intelligently,proactively and accurately.The intelligent risk prediction method for digital financial fraud aims to mine the rules and knowledge contained in the large-scale data to accurately and efficiently predict digital financial fraud risks.Firstly,to extract valuable information for digital financial fraud risk prediction from scenarios with large transaction volumes and low data density,the key technology of digital financial shallow feature selection is studied.Secondly,to solve the problem of unstable prediction performance of single classifiers in the unbalanced scenarios of digital financial fraud samples,the key technology of integrated prediction model construction is studied.Additionally,to analyze the important influence of historical transactions on current and future results,this thesis studies the key technology of deep temporal prediction model construction,excavates the complex temporal correlation in fraudulent transaction data.Finally,to explore the graph structure features between fraudulent transactions and transactions,and integrate the spatio-temporal characteristics of digital financial fraud,this thesis studies the key technology of deep spatio-temporal prediction model construction.The main research contents and innovation points are as follows:(1)Research on the feature selection method for digital financial fraud.Aiming at the redundant and irrelevant problems of digital financial fraud data,this proposes a shallow feature selection method that integrates improved glowworm swarm optimization,multi-fractal dimension(MFD),probit regression and artificial prior knowledge.First,MFD is employed as the evaluation criterion for the feature subset,and the improved glowworm swarm optimization is utilized as the search strategy to preliminarily select the feature subset of digital financial fraud risk.Combined with the financial background and digital financial fraud domain knowledge,probit is wielded to eliminate the factors that are not significantly related to digital financial fraud risk in the preliminary feature subset and combine them with the attributes selected by artificial prior knowledge to select the key features of digital financial fraud providing a high-quality database for subsequent predictions.(2)Research on the selective ensemble prediction method for digital financial fraud based on shallow features.Aiming at the defect of unstable prediction performance of single classification model commonly used in digital financial fraud prediction,a selective ensemble prediction approach combined information exchange glowworm swarm optimization(IEGSO)with difference measure is proposed.Firstly,a selective ensemble approach based on a double error measure is designed to reduce the computational resource consumption of ensemble learning.Wield the Bootstrap method to repeatedly extract the training set to obtain multiple base classifiers with large differences.Build a base classification pool,then calculate the double error measure of each base classifier,and sort by measure size in ascending order.Employing the majority voting,the base classifiers are accumulated and ensembled one by one according to the order until the ensemble accuracy is optimal,and the effectiveness of the method is theoretically analyzed.Secondly,in order to find the balance between the difference and the average precision between the base classifiers,a selective ensemble financial fraud prediction method that fuses the IEGSO and the difference measure is proposed.The base classification pool is pre-selected by the performance measure,and the remaining base classifiers after the pre-selection are re-selection by IEGSO,which is used to solve the practical problem of digital financial fraud risk prediction.(3)Research on the deep time series prediction of digital financial fraud risk.Aiming the complex time series feature selection of digital financial fraud risk,this thesis studies deep time series feature mining based on deep learning,and performs time series prediction for digital financial fraud risk.A digital financial fraud time series prediction method is designed based on the Dynamic Evolutionary Glowworm Swarm Optimization(DEGSO)and Long Short-Term Memory(LSTM)model.DEGSO is proposed by introducing the dynamic evolutionary mechanism and the directional mutation mechanism,which optimizes the search performance of GSO.Then,DEGSO is employed to optimize the main parameters of LSTM,and the best parameters combination of LSTM is obtained,which can effectively learn the temporal dependency in the digital financial fraud data and improve the accuracy of time series prediction.(4)Research on the spatiotemporal prediction methods for digital financial fraud risks.Combining the LSTM performed well in temporal feature extraction with the Graph Convolutional Neural Network(GCN)performed well in spatial feature extraction,a hybrid spatiotemporal prediction method for financial fraud based on GCN and LSTM is proposed.Thereinto,GCN is employed to extract the spatial dependency between different transactions,and then LSTM is wielded to mine the temporal dependency of financial transactions.Finally,deep spatial and temporal features contained in the digital financial fraud data are achieved,and the spatiotemporal prediction of fraud risk is carried out to improve the effectiveness and significance of the proposed model.This thesis systematically studies the key issues and key technologies in intelligent risk prediction for digital financial fraud.The research results have important theoretical significance for building an efficient and accurate digital financial fraud risk prediction model.At the same time provide financial practitioners with technology for investment decisions support is of great practical significance for maintaining financial stability and security.