Generating multi-type sequences of temporal events to improve fraud detection in game advertising.

摘要: Fraudulent activities related to online advertising can potentially harm the trust advertisers put in networks and sour gaming experience for users. Pay-Per-Click/Install (PPC/I) is one of main revenue models game monetization. Widespread use PPC/I model has led a rise click/install fraud events games. The majority traffic ad non-fraudulent, which imposes difficulties on machine learning based detection systems deal with highly skewed labels. From network standpoint, user are multi-type sequences temporal consisting event types corresponding time intervals. Time Long Short-Term Memory (Time-LSTM) cells have been proved effective modeling intrinsic hidden patterns non-uniform In this study, we propose using variant Time-LSTM combination modified version Sequence Generative Adversarial (SeqGAN)to generate artificial mimic fraudulent traffic. We also Critic instead Monte-Carlo (MC) roll-out training SeqGAN reduce computational costs. GAN-generated be used enhance classification ability event-based classifiers. Our extensive experiments synthetic data shown trained generator capability desired properties measured by multiple criteria.