Research On Multi-touch Attribution

In recent years,with the popularization and development of the Internet,online advertising has become the main profit method for Internet commercial applications and services such as social media,search engines,and e-commerce platforms.The goal of online advertising is to build information bridges between consumers and businesses through Internet products,promote products for businesses and acquire customers,and provide Internet users with product information that meets user needs.In the online advertising system,for advertisers,how to reasonably allocate to different media and channels with a fixed budget,and then use these channels to advertise to users,and ultimately obtain the maximum revenue,has always been a very important issue.Multi-touch attribution analysis brings opportunities to this demand.Multi-touch attribution analysis calculates the contribution of each touch point(ie,advertisement)to the user's conversion behavior in the user behavior path to further measure the influence of the media or channel to which the touch point belongs to conversion.To solve this problem,the existing work mostly proposes solutions based on the combination of sequence modeling and transformation prediction.However,this type of modeling has three problems.First,in business scenarios that lack the personal characteristics of users,traditional attribution models are difficult to perform effective attribution based on user preferences due to personalization issues.Second,in business scenarios with personal characteristics of users,it is difficult to fully learn about the interaction and long-term dependence between touch points for long user sequences.Third,for the periodic and non-periodic laws of the contacts in time,the model is also difficult to model the sequential mode.The above three problems make it more difficult for the model to accurately attribute touchpoints.In order to deal with the above three problems,this article proposes corresponding improvement methods:1.Aiming at the problem of personalized attribution,this paper proposes a new multi-task learning framework.In order to supplement the missing personal characteristics of users,this paper designs a multi-task learning framework.The framework includes click-through rate prediction tasks,conversion rate prediction tasks,and user channel preference prediction tasks.The channel preference prediction task uses causal inference theory to obtain a user interest representation vector through training,so that each given user behavior sequence can supplement the corresponding user-side features,and then perform personalized attribution to improve the accuracy of attribution sex.This paper verifies the effectiveness of the multi-task learning framework through experiments.2.Aiming at the interaction between the touch points in a longer sequence and the long-term dependent learning,this paper proposes a model based on the multi-head attention mechanism.For a given sequence of touch points,the previous method is difficult to capture the long-term dependence of the touch points in the sequence,and the calculation process cannot be parallelized,and the efficiency is low.With the help of the multi-head attention module,it can calculate the interaction and long-term dependence of each touch point and all other touch points in the sequence in parallel and bidirectionally,and use the information in the attribution calculation.This paper verifies the effectiveness of the multi-head attention model through experiments.3.Aiming at scenes with touch time characteristics,this paper proposes a time characterization module.Time information is very important for attribution,but previous work did not make effective use of it.In this regard,this paper designs a time representation module,which integrates time information into the vector representation,so that when the upstream model processes the contact sequence data,it can effectively use the information of the contact in the time series and pay attention to the touch points.Periodicity and non-periodicality between touch points and touch points can be used to obtain more accurate attributions.This paper verifies the feasibility and effectiveness of this method through experiments.