Research On The Method Of Identifying Online Fraudulent Transactions Based On Graph Analysis

In recent years,the development of scientific and technological innovation makes us into a new era of Internet Plus.With the continuous development of e-commerce,online shopping has become an indispensable part of all people's lives due to its convenience in all aspects.Incorrect online data caused by click farm can affect consumers' judgments.The click farm has strong anti-reconnaissance capability.For example,the platform will train the click farmers to simulate the real shopping process,making it more difficult for the algorithm to detect them.Another very important change is that many platforms tend to cultivate "two-sided buyers" through offering special discounts.Such buyers will do real shopping,and on the other hand,they will participate as part-time click farmers.Although the number of orders of them is not very large,they are very difficult to identify.Therefore,the identification of click farm has become one of the difficulties in the field of e-commerce.This thesis starts with the real shopping records of Tmall,and finds out how to identify this kind of highly collusive transaction behavior in view of the current pattern of fraudulent transactions.The main tasks as follow:1.This thesis proposes a method for identifying fraudulent transactions based on dense subgraphs mining.We propose to construct the network based on the user's purchase record.Due to the need of the click farm of e-commerce at a limited cost,it needs to improve its sales ranking as much as possible,so we redefine the weight of the products.According to the weight of the products,we use relevance between users.Since the behavior characteristics of different types of users are quite different,we use pruning based on the user's relevance to counter the pretending behavior of click farmers.After pruning,we can divide the network into several different subgraphs.Traverse each subgraph,based on the suspicious metric,then conduct dense sub-graph mining on each subgraph,and finally get dense subgraph.According to the subgraph,it is further confirmed that they are click farmers.Compared with the existing methods,when the results of the algorithm are basically the same as the current mainstream methods,the method in this thesis has greater advantages in large-scale deployment and utilization of server space.2.This thesis introduces ego network analysis into the field of anti-fraud in e-commerce.We analyze the users' ego networks,and find out the different behavior characteristics of different users.At the same time,because of the camouflage which is difficult to solve by traditional methods,based on the traditional ego network analysis,it is proposed to use the largest cluster with the most items of the ego network without ego and the feature network extracted by the sliding window to replace the traditional ego network to reduce disturbance caused by the purchase records of normal users and the camouflages of the click farmers.Using the features extracted from the feature networks,machine learning is used to find out the click farmers.The method based on ego network also has good results.At the same time,the interpretability of ego network analysis is also the key of this method.