Research On Mobile Advertising Fraud Detection Method Based On Graph Embedding

Mobile advertising,a new marketing method based on smart terminals,is more personalized,accurate,interactive and flexible compared to the traditional media platform.However,in pay-per-action model,malicious mobile ad publishers mimic legitimate users by artificially or programmatically methods to generate fraudulent actions on the advertisements to get more financial returns.This mobile advertising fraudulent behaviors poses a serious threat to the mobile advertising market.It is very challenging to identify fraudulent publishers since fraud technology evolves rapidly.Recently,advertising fraud detection has become a hot topic in the mobile advertising ecosystem.Graph-based approaches,which not only retain the natural structure of graphical data but also being more expressive and robust in representing the intricate data interactions,are widely applied to anomaly detection and fraud detection.Unfortunately,traditional graph-based approaches do not scale to large networks,and the popular deep learning approaches are incapable to handle graphical data without proper modifications.Recently,the graph embedding approaches have gain broad attentions.The intuition of such approaches is to learn the low-dimensional representations of the nodes to facilitate subsequent analysis.In this work,we propose a weighted heterogeneous graphical model(named by WMP2vec)to connect mobile users,application and advertisement.Each part in the graph is then embedded by a dense low-dimensional vector.In the meta-path random walk phase,the node transition probability is constrained by adding a weight offset factor to enhance the accuracy of the node representation.The experimental result in a real mobile advertising dataset shows that WMP2 vec outperforms several other graph embedding methods based on random walk.Furthermore,we explore the impacts of parameters in WMP2 vec algorithm.WMP2vec learns the relationship between different entities of mobile advertising ecosystem.However,the mobile advertising dataset contains other basic attribute information in addition to the graph structure information.In order to improve the accuracy of mobile ads fraud detection,we propose a hybrid fraud detection model based on graph embedding and deep learning(named by EHM),which takes both graph features in the heterogeneous graph and basic attributes of mobile applications into consideration with convolutional neural network.The model proposed in this paper is implemented in the deep learning framework named TensorFlow,and experimented in a real mobile advertising dataset.The experimental results show that EHM outperforms other popular supervised classification approaches on mobile ads fraud detection.Finally,we discuss the parameters in EHM.