Mobvista will attend ACM CIKM to share the latest research results in the field of anti cheating in mobile advertising

category:Internet
 Mobvista will attend ACM CIKM to share the latest research results in the field of anti cheating in mobile advertising


In recent years, with the rapid development of mobile Internet technology and the rise and popularization of intelligent mobile devices, the mobile advertising industry has developed rapidly. According to the emarketer report, the total global mobile advertising expenditure has reached 240.95 billion US dollars in 2019, and is expected to continue to grow to 286.5 billion US dollars in 2020. However, with the continuous growth of mobile advertising budget, mobile application installation cheating is becoming increasingly rampant, which not only causes advertising budget waste to advertisers, but also has a serious adverse impact on the advertising platform delivery effect and reputation.

Bots installation cheating refers to that cheaters use bots to simulate the application installation behavior of real users, so as to steal the advertising budget of advertisers. The reason why bots installation cheating is difficult to detect is that bots looks like a real user. It has real IP, device, etc., and can even carry out events within the application, such as opening the application at a specified time, and purchasing more.

At present, many mobile advertising fraud detection methods based on machine learning are aimed at solving the problem of advertising display or click cheating, but there is little research on the installation of cheating detection, especially the installation of bots cheating detection. According to Dr. Zhu Yadong, vice president of mobvista group, many of the current methods adopt integration methods and other technologies, such as randomforest and xgboost. Although these methods can mine rich cheating patterns through complex feature engineering, they cannot use structural information to build various relationships between entities, so these methods are not the best solution u3002

In this context, mobvista anti cheating team proposed a hybrid learning method, which combines graph neural network (GNN) and gradient promotion classifier. It can consider local context information and global context information at the same time, so as to better detect BOT installation cheating in mobile advertising. Instead of directly using the existing GNN model, the team constructed a heterogeneous graph according to the specific business of mobvista and the characteristics of BOT installation cheating, and designed a novel message passing mechanism to extract local context information; and a pre trained gradient promotion classifier model was used to extract global context information. The architecture of the botspot model is shown in the following figure:

The results show that botvista has more advantages in the detection of the data of botbotspot by installing the model. When the precision is 90%, the recall rate of botspot model is improved by at least 2.2% and 5.75% respectively compared with all other baseline methods on two offline datasets.