Anomaly Detection Of Intelligent Securities Trading Behavior Based On Mixed Model

For a long time,financial fraud detection has been a research hotspot in the field of anomaly detection.Most financial fraud detection models only focus on the anomaly of single user transaction behavior,such as credit card,insurance and other fields; Some fraud detection models consider sequence information,but only detect whether there is fraud at a certain time point in the sequence,such as capital chain,transaction flow and other scenarios.In this paper,we face the problem of anomaly detection in multiple scenarios under securities,where the time point of fraud is often uncertain and there are joint anomalies of multiple behavior sequences and single behavior anomaly at the same time;In addition,real securities scenario datasets often only contain user level label data,and there is no label on the behavior or time point level.In view of these unique challenges in the domestic securities field,the major contributions as follows:(1)Firstly,face the complexity and professionalism of securities trading scenarios,this paper proposes a hybrid model of transaction behavior anomaly detection based on multi-instance learning.The model decision tree part provides domain expert features and strong interpretability.In the deep learning part,this paper introduces multi-instance learning into financial fraud detection for the first time to solve the special semi supervision problem in the securities scene.By using the relationship between the instance and the package,the similarity matrix is given through the similarity calculation formula designed in this paper,so that each package is given a new representation,and support vector machine is used for feature selection.In addition,the model designs different weights for positive and negative samples in hinge loss to alleviate the problem of sample imbalance,and the characteristics of multi-instance learning also add interpretability to the deep learning part.Finally,the effectiveness of the proposed method is proved by experiments.(2)Secondly,a few of the hybrid models proposed by predecessors have the above shortcomings,but also have high computing costs,which is fatal for securities fraud detection scenarios.Therefore,this paper further attempts to balance the online challenges of effectiveness,performance and interpretability from the perspective of landing applications,and proposes a hybrid model for anomaly detection of securities trading behavior based on attention mechanism.The machine learning part of the model is the same as that of the previous model; In the deep learning part,the model obtains the representation of different user behavior sequences through the sequence model,in which two attention mechanisms are also used to fuse the representation and internal information between the behavior sequences,respectively,to capture the specific sequence and sequence interval causing the exception.Finally,the effectiveness of the proposed method is proved by experiments.(3)Thirdly,in view of the lack of one-stop securities scene fraud detection system on the market,this paper develops an intelligent transaction security evaluation system based on the designed model,including automatically generating visual user transaction behavior detection reports,providing automated scripts for the use of securities platforms,etc.,and has been deployed in Shanghai Huaxin Securities Company and its subsidiaries in relevant business scenarios for nearly a year,Good real benefits have been achieved.