A Probabilistic Multi- Touch Attribution Model for Online - - PowerPoint PPT Presentation

A Probabilistic Multi- Touch Attribution Model for Online Advertising

Author : Wendi Ji, Xiaoling Wang, Dell Zhang Source : CIKM’ 16 Advisor : Jia-Ling Koh Speaker : Chia-Yi Huang Date : 2017/02/21

Outline

▸ Introduction

▸ Method ▸ Experiment ▸ Conclusion

2

Introduction

3

Introduction

4

▸ Probabilistic Multi-Touch Attribute

▸ Whether a user will convert ▸ When she will convert

Introduction

5

▸ Survival Analysis

▸ Survival function(⽣甠存函數)：


S(t) = Pr(T>t), T為⽣甠存時間, t為某個時間

▸ Lifetime distribution function(衍⽣甠函數)：


F(t) = 1 - S(t) = Pr(T <= t)

▸ Hazard function(危險函數)：


λ(t) = F’(t) / S(t)

Outline

▸ Introduction

▸ Method

▸ Experiment ▸ Conclusion

6

7

▸ Weibull Distribution

▸ When α < 1, the hazard rate is a monotonic

decreasing function.

▸ When α = 1, the hazard rate is a constant over

time 1/λ

▸ When α > 1, the hazard rate is a monotonic

increasing function

Method

8

▸ Weibull Distribution

Method

9

▸ Probabilistic Model

▸ User => {1,…,U} ▸ Advertising Channels => {1,…,K} ▸ Browsing path bu of user u 


=>

▸ lu : length of the ad browsing path bu ▸ : set of features

Method

whether conversion occurred when conversion occurred

10

▸ Probabilistic Model

▸ : advertising channel ▸ : timestamp of click / impression ▸ : 1 is conversion has occurred , 0 o.t.w ▸ : last timestamp of the observation window ▸ : = 1, timestamp of the conversion occurred ▸ : 1 is conversion will happen , 0 o.t.w ▸ : conversion delay of an ad exposure

▸ : the elapsed time

Method

11

▸ Probabilistic Model

▸ When Y = 1

Method

12

▸ Probabilistic Model

▸ When Y = 1

Method

Train

13

▸ Probabilistic Model

▸ When Y = 1

Method

Train

14

▸ Probabilistic Model

▸ When Y = 0

Method

15

▸ Probabilistic Model

▸ When Y = 0 


▸ Y = 1 + Y = 0

Method

16

▸ Parameter Estimation ▸ Multi-Touch Attribution

Method

17

▸ Conversion Prediction

Method

Outline

▸ Introduction ▸ Method

▸ Experiment

▸ Conclusion

18

19

▸ DataSet

▸ A large real-world dataset provided by Miaozhen, a

leading marketing company in China.

▸ The timestamp, the user ID, the channel ID, the

advertising form, the website address, the type of

• peration system and browser, etc.

▸ 1.24 billion data records ▸ 59 million users(0.01% convert) and 1044 conversions

available.

▸ Involved 2498 channels with 40 forms (e.g. iFocus, Button,

Social Ad) and 72 websites (e.g. video websites, search engines, social networks)

Experiment

20

▸ Baseline

▸ AdditiveHazard ▸ Simple Probability ▸ Time-aware ▸ Logistic Regression

Experiment

21

▸ Conversion Prediction

Experiment

22

▸ Conversion Prediction

Experiment

included

23

▸ Conversion Prediction

Experiment

24

▸ Attribution Analysis

Experiment

25

▸ Attribution Analysis

Experiment

Outline

▸ Introduction ▸ Method ▸ Experiment

▸ Conclusion

26

Conclusion

▸ The PMTA model for conversion attribution which

takes into account both the intrinsic conversion rate of a user and the conversion delay.

▸ The PMTA model can be applied to conversion

prediction.

▸ The PMTA model is fitted to the observed data

(conversion rate and conversion delay) rather than relying on simplistic assumptions.

▸ The PMTA model has been evaluated on a large

real-world dataset.

27

THANK YOU