Research And Implementation Of Intelligent Optimization Technology For E-commerce Advertising Content Based On Multi-task Deep Learning

Nowadays more and more people have got used to purchase on E-Commerce platforms,such as Amazon,Ebay,Taobao and Jingdong.In these platforms,sponsored search(SS)is a standard mechanism for promoting the sales revenue of online merchants,as well as bringing benefits to the platform.Sponsored search is also a keyword-based advertisement based on search.Compared with traditional advertisements,sponsored search has the advantages of strong purpose,strong pertinence,quick effect,low cost,etc..In short,it has obvious advantages.Most of the previous research on SS was focused on bidding optimization and click prediction(in general search engines).The thesis studies a novel problem of SS for E-Commerce platforms: how to attract query users to click product advertisements(ads)by intelligently optimizing the advertising content.In other words,presenting them features of products that attract them.This not only benefits merchants and the platform,but also improves user experience.The problem is challenging due to the following reasons:(1)the ad content needs to be carefully manipulated so as to effectively attract users' attention without affecting search experience and damage to the merchants' business revenue.(2)It is difficult to obtain users' explicit feedback of their preference in product features.(3)With the rapid growth of mobile Internet,a great portion of the search traffic in E-Commerce platforms is from their mobile apps(e.g.,nearly 90% in Taobao).The situation would get worse in the mobile setting due to limited space.The thesis is focused on the mobile setting and propose intelligent optimization of advertising content.That means manipulating ads' titles by adding a few selling point keywords(SPs)to attract query users.This thesis models it as a personalized attractive SP prediction problem.The contributions include:(1)this thesis explores various exhibition schemes of SPs in ads and analyze their effectiveness.(2)this thesis proposes a surrogate of user explicit feedback for SPs preference and further develop a multi-task model which employs click-through rate(CTR)prediction as an auxiliary task to boost the performance of the SPs prediction task.(3)For the multi-task learning model,this thesis designs an offline experiment to explore the effectiveness of additional features about users and queries,and the impact of different clicks of users on the model under multiple tasks.This thesis carries out both large-scale offline evaluation and online A/B tests in Taobao.The thesis uses the abundant user click log data on Taobao app to conduct experiments on multiple models to illustrate the effectiveness and superiority of the multi-task model proposed by this thesis.It should be pointed out that a variant of the best model has already been deployed in Taobao,leading to a 2% increase in revenue per thousand impressions and an opt-out rate of merchants less than 4%.