Research On Fraud Detection Algorithms For Mobile Advertising

Advertising promotion based on mobile terminal devices is one of the most important advertising marketing channels today.In addition to mobile browser websites,application software on mobile devices is also an important carrier of mobile advertising.Providing advertising fields and user traffic for advertisers is an important source of income for many mobile applications.However,in order to obtain more profits,some mobile applications will fool more advertising revenue by forging fake user traffic,resulting in serious loss of advertising promotion funds for advertisers.How to identify fraudulent advertising traffic is an important challenge for advertisers and advertising platforms.The traditional mobile advertising fraud detection scheme mainly relies on manual rules formulated by data researchers.With the continuous updating of fraud devices and fraud methods,the original manual rules have become increasingly difficult to distinguish fraudulent user traffic.Therefore,this paper proposes a mobile advertising fraud detection scheme that uses a combination of multiple machine learning algorithms for fraud detection in mobile advertising scenarios.The main innovative work of the program is as follows:(1)In view of the high recognition rate of the conditional discrimination rules in the fraud detection scenario and the existence of a large number of sparse category features in this scenario,this paper constructs the GBDT model and the neural network model separately,by using two different single models combined way to improve the comprehensive detection effect of the program.(2)In this scenario,there are a large number of new users without historical data,but the user-media relationship network can reflect the fraud tendency of these new users.In view of these characteristics,this paper uses graph embedding technology to construct user's relationship vector.At the same time,in order to make the downstream detection model fully mine the user relationship information in the embedded vector of the graph,this paper combines the characteristics of the wide & deep neural network model to be good at mining sparse category features,and proposes an improved model Emb Deep Fm that supports the input of embedded vectors.Experiments show that the neural network model combined with graph embedded vector information can effectively improve the fraud detection ability of new users without historical data.(3)In view of the characteristics of GBDT model and neural network model that are good at predicting samples with different feature distributions,this paper proposes an integrated learning combination strategy DG-Blending based on the combination of difference grouping,which makes full use of the differences in the capabilities of the two models.Through experimental comparison,the combination effect of the new model combination strategy is significantly improved compared with the existing combination strategy.Experimental results show that the mobile advertising fraud detection scheme proposed in this paper has a very prominent fraud detection effect under the AUC indicator.This solution can well identify complex patterns of advertising click fraud,and provides a feasible solution for the application of machine learning algorithms in the field of mobile advertising fraud detection.