Research On Anti Fraud Of Online Loan Based On Spatiotemporal Correlation

Online loan develops convenient loan business through the Internet,meeting the loan needs of different groups of people.With the development of online loan business,online loan fraud has also become more and more fierce.Unlike traditional loan data,which has a strong correlation,online loan data show weak correlation.This makes it difficult to model online loan anti-fraud.Different from the common research perspective of online lending,this paper studies the characteristics of bad debts that are different from normal loans in time and space,and finds that the potential fraud in bad debts is more related in time and space.By modeling the spatiotemporal connection in the process of online lending to improve the risk control system,we can prevent the lending of spatiotemporal abnormal online lending and avoid economic losses.Main tasks as follows:(1)Based on the spatiotemporal analysis of online loan data,this paper finds out the differences between bad loans and normal loans in terms of operation cycle,spatiotemporal diversity,urban cohesion and spatiotemporal outbreak.For the provinces,cities,longitudes and latitudes of online loans,this paper uses the sliding window to count the number of online loans under different time windows of these geographical factors,so as to establish the spatiotemporal connection features of online loans.In this paper,the features of the basic model are composed of the original features and their derivative features,combined with the spatiotemporal connection features established by the sliding window,and combined with the Light GBM model to predict the online loan fraud.(2)In this paper,two novel modeling methods are proposed for the spatiotemporal correlation of online loans based on Convolutional Neural Networks.Different from the geographic information of online loans,we simply use the sliding window to extract the spatiotemporal statistical values.The first one is to build a two-dimensional matrix of time and space for observing online loans and neighboring online loans.The second method is to form a spatial matrix of temporal trend values by learning the temporal trend values of the Long Short-Term Memory(LSTM)of the observed loans and neighboring loans.Finally,this paper uses convolution neural network(CNN)to extract the spatiotemporal correlation feature vector of observation loan as spatiotemporal correlation feature.Compared with simply using sliding window,the prediction effect is better in KS value and AUC value.(3)In this paper,an innovate relative neighbor index is proposed.Different from observing the spatiotemporal connection of online loans from a fixed spatial perspective,this paper limits the number of fixed nearest neighbor online loans,and calculates the spatial location of the nearest neighbor online loans by convex hull to dynamically determine the neighborhood domain.KN relative nearest neighbor index is established by the ratio of aggregation measures of large and small neighborhood domains in a certain time window.In addition,this paper uses the seq2seq(sequence to sequence)model based on LSTM to extract the feature vector between sequences as the spatiotemporal correlation feature from the KN relative nearest neighbor index under different time windows,compared with the basic model,the simply using of sliding window,this method also has better performance.