Probabilistic Generative Models For User Behavior Prediction

In recent years, with the rapid growth of Internet, the Internet services such as large online social network and online advertising is getting more and more popular.Taking viral marketing as an example, in order to improve the e?ciency and maximize the ROI(Return of Investment) of a marketing campaign, it is important to know how to model user behavior and identify the most in?uential users to target. For example, in online advertising, analyzing the impact of different ad channels on users’ decision of buying products and reallocate ad budget accordingly is important to improve the return on investment(ROI). Massive amount of user behavior data is generated by the rapid growing Internet services, which offers us a great opportunity to model and analyze users’ behavior. This work mainly focus on how to model, analyze and predict users’ behavior.In this work, we mainly focus on how to model users’ behavior under two different settings, including online social netowrk and online advertising.Modeling the information diffusion process over social networks and the iteraction between users and information has attracted a lot of research attention. In particular, information diffusion cascades can be useful for not only inferring the underlying structure of the in?uence network, but also providing insights on the properties of information itself. We address the problem of jointly modeling the in?uence structure and the hotness of the information itself based on the temporal events describing the process of the information diffusion. Specifically, we extend the multi-dimensional Hawkes process, which captures the mutual-excitation nature of information cascading, to further incorporate the hotness of the information being propagated. In the proposed method, the hotness of information and the network structure are modeled in a unified and principled manner, which enables them to reinforce each other and thus enhances the estimation of both. Experiments on both real and synthetic data show that our algorithm typically outperforms several existing methods and accurately estimates the hotness of information from the observed data.How to model and predict users’ purchasing behavior and distribute the credit to all related advertisements based on their corresponding contributions has recently become an important research topic in online advertising both in industry and academia.Compared to traditional advertising, such as TV and print advertising, online advertising generates massive amount of trackable user data, upon which we can build models to predict user behavior and attribute ad channels. Traditional rule-based models have been widely used in industrial practice. The drawback of such rule-based models lies in the fact that the rules are not derived form the data but only based on simple intuition.With the ever enhanced capability to tracking advertisement and users’ interaction with the advertisement, data-driven models become an important research direction. We propose a new data-driven model based on survival theory. By adopting a probabilistic framework, one key advantage of the proposed model is that it is able to remove the presentation biases inherit to most of the other models. In addition to predict user’s‘conversion’ probability, the proposed model is also able to distribute the attribution of different ad channels. We validate the proposed method with a real-world data set obtained from a operational commercial advertising monitoring company. Experiment results have shown that the proposed method is quite promising in both conversion prediction and attribution.