Time-Aware Conversion Prediction And Attribution Analysis

With the rapid development of data storage and data analysis,E-commerce is becoming the leading media for more and more companies to promote their produces and services.Com-putational advertising is a most vital component of E-commerce,of which targeting techniques make personalized adverting possible.Based on the context and feedback information of users,online advertising systems are able to deliver ads to the user with the highest possibility to respond.Click-through rate and conversion rate are two main criterion of the advertising per-formance utilized by the majority of advertisers.A higher click-through rate or conversion rate indicates that an advertising champaign attracts more users for the advertiser or,we could say,this advertising champaign is more effective.In E-commerce systems,conversion rate is the most significant metric to measure the performance of advertising and recommending.It refers to the proportion of users who take a specified action,e.g.purchases,registrations or searches.Here,the term "conversion" refers to the action of transforming a visiter to an actual consumer.Compared with click-through rate,there are two technical challenges for conversion rate prediction:(1)how to predict conversion time and(2)how to attribute the credits of different channels for the conversion.The conversion rate of a user is closely related to time.For ex-ample,recommending television to a user who has just bought one might likely not bring in another purchase,because user will not buy an extra big-ticket item of the same kind within a short period.To deliver the right advertisement at a wrong time leads to a waste of the op-portunity to influence users.In actual advertising systems,the most advertisers are primarily interested in which users will convert in short,specially in re-targeting advertising.On the other hand,an online advertising champaign always launches through multiple advertising channels,e.g.display ads,paid search ads,social ads and etc.A user is influenced by lots of ads and her/his conversion is the result of all relative ads.Therefore,a reliable attribution strategy is required to measure the effects of different channels to the user conversions,which is helpful for advertisers to optimize their champaign.In this paper,we focus on the two challenges we mentioned above and the major contributions are as follows:(1)A new research problem that how to advertise for a given period(e.g.three days or a week)is proposed.We build a personalized lifetime model based on Weibull distribution to describe the conversion time.To make the proposed model more comprehensive,we build it under Bayesian framework and estimate the parameters with EM algorithm.In order to predict the behaviors of a user within a given period,we propose a rank-based time-aware conversion prediction model.We evaluated the proposed model on two real-world datasets.The experi-mental results demonstrate that our lifetime model fit the conversion times well and illustrate its effectiveness in time-aware conversion prediction.(2)A novel data-driven attribution analysis model is proposed for conversion attribu-tion,named probabilistic multi-touch attribution model.Inspired by survival analysis,we use Weibull distribution to model the conversion delay between ad exposures and conversion time,which indicates the effects of different advertising channels changing with time.Drawing the analogy between conversions and the deaths in survival analysis,we quantitatively model the contribution of a channel with hazard rate.In addition,the proposed attribution model is built in the probabilistic framework.It is also applied to the task of conversion prediction,and takes both the user’s intrinsic conversion rate and conversion time into consideration.(3)An additional multi-touch attribution model is also proposed for conversion attribution.Borrowing the concepts of point process,we assume the influence of an ad exposure fades with time and the effects of all relative ads on the path to conversion are additive.We directly use hazard rate in survival analysis to model the influence of an ad exposure to the conversion,of which the value is determined by the influence strength and its decaying speed.According to the basic assumptions,the distribution of conversion time is calculated by the sum of hazard rates of all relative ads.Furthermore,a conversion rate prediction model is also proposed based on the additional multi-touch attribution,which considers both whether the user will convert and when the conversion will occur.We evaluated the performance of the proposed attribution model on a real-world dataset.The experimental results demonstrate its effectiveness in both conversion attribution and conversion prediction.In conclusion,this paper aims at two challenges of conversion prediction and proposes a series of strategies of modeling conversions based on survival analysis and point process,including a time-aware conversion prediction model and two multi-touch attribution models.The researches in this paper compose a a relatively complete system,which are consistent and sustainable.We carry on lots of theoretical analysis and experiments on real-world dataset.The results illustrate that our research strategies and three proposed models have good effects on the two challenges of conversion prediction.