Evaluating Impact Of Media Communication Based On Temporal Point Process

With the development of science and technology,the form of the media has undergone tremendous changes,from print publications to radio,television,and now to the Internet.In-ternet technology makes it possible to collect the large-scale data of users' interaction with the media.This has led to the need for research of data-driven media communication impact,namely,designing model or method to evaluate the impact of media from the data.Focusing on the advertising data,this paper studies impact evaluation of digital advertising channel.The ads are served through multiple channels on the media.The effect of advertising we observed is a cumulation of multiple advertising channels.How to separate the effect of a single channel or media,respectively,is a challenge.We model the behavior of users based on temporal point process.Building on the top of the additive hazard model of survival theory,we extend in two aspects,1)introducing a factor to model the interaction between advertising chan-nels,and 2)modeling multiple purchase behaviors of a user.Two data-driven digital advertising attribution models are designed to calculate the influence of each advertising channel.Most of the existing data-driven attribution models fail to take account of the interaction between ads.We use the concept of vector synthesis in the attribution model to introduce the two factors synergy and antagonism,in order to achieve non-linear cumulation of multiple advertis-ing channel s influence.Experiments on real advertising dataset show that the prediction of user behavior becomes more accurate after considering the interaction.It was found that the antago-nism between advertising channels was significantly greater than synergy,especially when the ads are of the same type or the same website,while the ads of different websites showed more independence.In order to solve the problem that the existing model based on the survival theory can not deal with the user s multiple purchase,we model the purchase behaviors as time points and use the non-homogeneous Poisson Point Process to model the user s purchase behaviors.The conditional intensity function of the Poisson Point Process can be used to model the erfect of multiple advertising channels on the user s purchasing behavior as the user interacts with the advertising media.We use the MM algorithm to optimize the model iteratively,which can be very efficient to learn the parameters of the model.Experiments show that,by taking more the user s purchase information into account,the user s conversion prediction becomes more accurate.The model of interaction between user and advertising media is not the ultimate purpose of this paper.After fitting the model parameters with the training data,this paper calculates the contribution of each advertising channel,and combines the advertising websites and types to get the media evaluation of communication effects.Comparing the results of the two models proposed in this paper with those of other benchmark models,we find that the two models proposed in this paper are similar to each other in ranking of results of the advertising media,and they are more reasonable than other models.Their evaluation results,one of which is Search Engine is the most effective website,are also more in line with industry awareness.According to the comparison of evaluation results on types and websites,the difference among influence of different websites is bigger than that of different types.