Research On Mobile Advertising Fraud Detection And Recommendation Models

With the prosperity of miscellaneous types of terminal intelligent devices(mobile phones,tablet,PCP,etc.),the internet advertisement market is moving from PCs to mobile terminal devices.Mobile advertising is a novel fashion of marketing that relies on terminal devices.Two of the current,primary concerns in the context of mobile advertising information services are,namely,mobile advertising fraud detection and advertising recommendation.To detect the fraudulent apps for mobile advertising service is difficult,since fraudulent traffic often mimics behaviors of legitimate users and evolves rapidly.In this paper,we propose a novel bipartite graph-based propagation approach,iBGP,to detect advertising fraud built in malicious mobile apps.We first exploit the characteristics of mobile advertising users' behavior and identify two persistent patterns: power law distribution and pertinence,and formulate both concepts to construct an initial score learning algorithm to compute the initial fraud scores of non-seed nodes.We propose a weighted graph propagation algorithm to propagate the scores of all nodes in the user-app bipartite graphs until convergence.Experiment results in real-life mobile datasets demonstrate that iBGP outperforms other popular graph-based propagation methods.For the matter of mobile advertising recommendation,we study two sub-problems under the Collaborative Filtering framework,namely implicit feedback quantification and top- recommendation under neighborhood-based model.With respect to the issue of implicit feedback quantification,most previous methods intuitively assign the implicit feedback with binary scores or numerical scores,which might underfit or even misrepresent the feedback especially for hybrid and complex recommender systems.In this paper,we propose a implicit feedback quantification model that identify the score of the feedback automatically for each user based on ones observed implicit feedback list.Our model consists of two parts: sorting and scoring.To evaluate the feasibility of our model,we conduct experiments on Spark platform using real-life mobile advertising dataset.Results showcase that our model consistently outperforms the competing methods on both neighborhood-based models and latent factor models.For the issue of large-scale top- recommender system under neighborhood-based model,we propose a two-staged user similarity based top- recommendation approach(MobRec),which combines user clustering and top- ranking into one recommender system for large mobile in-app advertising.MobRec uses the mobile user feature classification and similarity aggregation to partition the large user-user similarity computing into an offline clustering phase and an online nearest neighbor computing phase.Based on the nearest neighbors,a novel preference model is proposed to provide accurate top- recommendation,which is refined for the mobile ads scenario.MobRec is validated over a large real-world mobile advertising dataset on the Spark platform;experimental results show that MobRec outperforms the popular methods as well as several state-of-the-art methods.