Advanced Attribution Model, 
Analytics Summit, November 13th , 2014 
Aspa Lekka 
a.lekka@foodpanda.com
Foodpanda is a global online food delivery marketplace 
Company: 
 www.foodpanda.com 
 Founded in 2012 
 Present in more than 45 countries 
Business Intelligence & Analytics: 
 Google Analytics and Google Analytics Premium 
 Team of 8 people
TV 
Mobile 
App Ad 
SEM Conversion 
Customer Journey
Multiple channels: 
•SEM 
•Display 
•CRM 
•Affiliate 
•Radio 
•Price Comparison 
•….. 
Marketing budget allocation 
Get new visitors 
AWARENESS 
Maximize orders 
ACTION 
0 50 100 150 200 250 
Social 
Direct 
Display 
Affiliate 
Price comparison 
Search 
Last click- orders TOTAL: 503 
0 20 40 60 80 100 120 140 160 180 
Social 
Direct 
Display 
Affiliate 
Price comparison 
Search 
Last click - orders TOTAL: 438 
Awareness 
Interest 
Desire 
Action 
Why attribution model?
Every channel gets attributed the correct orders
Calculating the CAC for Marketing Campaigns 
NEW CUSTOMERS ACQUIRED 
Campaign 1 BUDGET Monday Tuesday Wednesday Thursday Friday Saturday Sunday TOTAL 
Monday 1,000 € 13 13 
Tuesday 1,100 € 21 21 
Wednesday 900 € 22 22 
Thuerday 1,000 € 25 25 
Friday 850 € 23 23 
Saturday 1,200 € 24 24 
Sunday 1,000 € 26 26 
CAC 
77 € 
52 € 
41 € 
40 € 
37 € 
50 € 
38 €
How efficient was the Marketing Budget? 
NEW CUSTOMERS ACQUIRED 
Campaign 1 BUDGET Monday Tuesday Wednesday Thursday Friday Saturday Sunday Monday Tuesday Wednesday Thursday Friday Saturday TOTAL 
Monday 1,000 € 10.5 2.9 14.5 10.8 12.3 19.0 9.0 79.0 
Tuesday 1,100 € 10.5 6.6 3.9 13.3 3.1 6.2 13.9 57.4 
Wednesday 900 € 5.2 4.1 8.0 9.4 8.1 18.1 17.8 70.6 
Thuerday 1,000 € 12.7 20.2 17.3 5.8 3.9 4.0 9.7 73.6 
Friday 850 € 4.7 7.8 12.0 6.6 6.0 20.5 2.6 60.2 
Saturday 1,200 € 19.4 16.7 7.6 17.6 19.0 6.1 16.1 102.4 
Sunday 1,000 € 7.4 2.2 5.9 15.2 13.4 16.4 11.7 72.1 
CAC* 
13 € 
17 € 
14 € 
14 € 
17 € 
10 € 
14 €
Data Driven Models for Attribution 
1. SHAPLEY VALUE 
2. SURVIVAL ANALYSIS 
PATH ANALYSIS 
NETWORKS 
STRUCTURAL EQUATION MODELING 3.
The Shapley Value 
SHAPLEY VALUE
The Shapley Value 
The Shapley value is a way to assign credit among a group of “players” who cooperate for a certain end 
An example: 
• 3 players 
(2 with right glove and 1 with left glove) 
• Goal: Form a pair 
• Assign credit to each player after forming a pair 
There are two possible pairs that we can form and in both of them, Player 1 needs to be involved. Therefore, Player 1, is of more importance compared to Player 2 or 
Player 3. Consequently, when sharing the profits, he should get a bigger part compared to Player 2 (if we case 1 is true) or Player 3 (if case 2 is true)
• 3 channels 
• A click chain that consists of these 3 channels and led to 500 transactions 
• Evaluate the contribution of each channel to these 500 transactions 
100 
125 
50 
270 
375 
350 
500 
The Shapley Value
SEM 
100 
DISPLAY 
270-100=170 
SEO 
100 500-270=230 
125 
50 
270 
375 
350 
500 
SEM 
100 
SEO 
375-100=275 
DISPLAY 
500-375=125 
DISPLAY 
125 
SEM 
270-125=145 
SEO 
500-270=230 
DISPLAY 
125 
SEO 
350-125=225 
SEM 
500-350=150 
SEO 
50 
SEM 
375-50=325 
DISPLAY 
500-375=125 
SEO 
50 
DISPLAY 
350-50=300 
SEM 
500-350=150 
Calculating the Shapley Value
Calculating the Shapley Value 
SEM 
100 
DISPLAY 
270-100=170 
SEO 
500-270=230 
SEM 
100 
SEO 
375-100=275 
DISPLAY 
500-375=125 
DISPLAY 
125 
SEM 
270-125=145 
SEO 
500-270=230 
DISPLAY 
125 
SEO 
350-125=225 
SEM 
500-350=150 
SEO 
50 
SEM 
375-50=325 
DISPLAY 
500-375=125 
SEO 
50 
DISPLAY 
350-50=300 
SEM 
500-350=150 
SEM’s expected marginal contribution is: 
DISPLAY’s expected marginal contribution is: 
SEO’s expected marginal contribution is: 
SEM 
162 
(0.32) 
DISPLAY 
162 
(0.32) 
SEO 
176 
(0.36) 
500 orders attributed to the three channels :
Survival Analysis 
SURVIVAL ANALYSIS
Target 
population 
Treatment 1 
Dead 
Alive 
Treatment 2 
Dead 
Alive 
Event 
Event 
TIME 
What is the patient’s 
probability to be still 
alive after 20 years? 
Survival Analysis
Visitors 
Channel 1 
Event 
Conversion 
NO 
Conversion 
Channel 2 
Conversion 
NO 
Conversion 
Event 
TIME 
What is the visitors’ 
probability to convert 
after 30 days or 5 visits? 
Survival Analysis
Day1 
Day2 Day5 Day2 
Day3 Day7 Day8 
DAY 1 / VISIT 1 
DAY 3 / VISIT 3 
0 1 0 1 
DAY 6 / VISIT 6 
0 1 
DAY 9 / VISIT 9 
0 1 
Survival Analysis
VISITORS 
TIME 
START END 
: censored observation 
: event (conversion) 
Censored observation: 
There is not “time to 
event” recorded because: 
•Loss of follow up 
 Drop out 
 Conversion due to a 
cause that is out of 
our interest 
•End of the study 
Survival Analysis
Survival Analysis 
Estimate time-to-event for a group of 
individuals, such as time until a visitor purchases 
To compare time-to-event between two or 
more groups, such as visitors that have clicked 
on a Display ad compared to visitors that have 
not clicked on a Display ad. 
To assess the relationship of co-variables to 
time-to-event, such as: does number of clicks, 
pages viewed, or time on site effect the decision 
to purchase?
DATA DRIVEN MODELS 
ADJUSTED TO EACH CASE 
LIMITED TUTORIALS 
ADJUST FORMULAS TO 
YOUR DATA 
(QUITE) EASY TO SET UP 
LINK ADS WITH ONSITE BEHAVOR 
FIND COST EFFICIENT CLICK CHAINS 
MERGE OFFLINE AND ONLINE DATA 
Evaluation & Suggestions
Thank you! 
Aspa Lekka 
a.lekka@foodpanda.com

Advanced attribution model

  • 1.
    Advanced Attribution Model, Analytics Summit, November 13th , 2014 Aspa Lekka a.lekka@foodpanda.com
  • 2.
    Foodpanda is aglobal online food delivery marketplace Company:  www.foodpanda.com  Founded in 2012  Present in more than 45 countries Business Intelligence & Analytics:  Google Analytics and Google Analytics Premium  Team of 8 people
  • 3.
    TV Mobile AppAd SEM Conversion Customer Journey
  • 4.
    Multiple channels: •SEM •Display •CRM •Affiliate •Radio •Price Comparison •….. Marketing budget allocation Get new visitors AWARENESS Maximize orders ACTION 0 50 100 150 200 250 Social Direct Display Affiliate Price comparison Search Last click- orders TOTAL: 503 0 20 40 60 80 100 120 140 160 180 Social Direct Display Affiliate Price comparison Search Last click - orders TOTAL: 438 Awareness Interest Desire Action Why attribution model?
  • 5.
    Every channel getsattributed the correct orders
  • 6.
    Calculating the CACfor Marketing Campaigns NEW CUSTOMERS ACQUIRED Campaign 1 BUDGET Monday Tuesday Wednesday Thursday Friday Saturday Sunday TOTAL Monday 1,000 € 13 13 Tuesday 1,100 € 21 21 Wednesday 900 € 22 22 Thuerday 1,000 € 25 25 Friday 850 € 23 23 Saturday 1,200 € 24 24 Sunday 1,000 € 26 26 CAC 77 € 52 € 41 € 40 € 37 € 50 € 38 €
  • 7.
    How efficient wasthe Marketing Budget? NEW CUSTOMERS ACQUIRED Campaign 1 BUDGET Monday Tuesday Wednesday Thursday Friday Saturday Sunday Monday Tuesday Wednesday Thursday Friday Saturday TOTAL Monday 1,000 € 10.5 2.9 14.5 10.8 12.3 19.0 9.0 79.0 Tuesday 1,100 € 10.5 6.6 3.9 13.3 3.1 6.2 13.9 57.4 Wednesday 900 € 5.2 4.1 8.0 9.4 8.1 18.1 17.8 70.6 Thuerday 1,000 € 12.7 20.2 17.3 5.8 3.9 4.0 9.7 73.6 Friday 850 € 4.7 7.8 12.0 6.6 6.0 20.5 2.6 60.2 Saturday 1,200 € 19.4 16.7 7.6 17.6 19.0 6.1 16.1 102.4 Sunday 1,000 € 7.4 2.2 5.9 15.2 13.4 16.4 11.7 72.1 CAC* 13 € 17 € 14 € 14 € 17 € 10 € 14 €
  • 8.
    Data Driven Modelsfor Attribution 1. SHAPLEY VALUE 2. SURVIVAL ANALYSIS PATH ANALYSIS NETWORKS STRUCTURAL EQUATION MODELING 3.
  • 9.
    The Shapley Value SHAPLEY VALUE
  • 10.
    The Shapley Value The Shapley value is a way to assign credit among a group of “players” who cooperate for a certain end An example: • 3 players (2 with right glove and 1 with left glove) • Goal: Form a pair • Assign credit to each player after forming a pair There are two possible pairs that we can form and in both of them, Player 1 needs to be involved. Therefore, Player 1, is of more importance compared to Player 2 or Player 3. Consequently, when sharing the profits, he should get a bigger part compared to Player 2 (if we case 1 is true) or Player 3 (if case 2 is true)
  • 11.
    • 3 channels • A click chain that consists of these 3 channels and led to 500 transactions • Evaluate the contribution of each channel to these 500 transactions 100 125 50 270 375 350 500 The Shapley Value
  • 12.
    SEM 100 DISPLAY 270-100=170 SEO 100 500-270=230 125 50 270 375 350 500 SEM 100 SEO 375-100=275 DISPLAY 500-375=125 DISPLAY 125 SEM 270-125=145 SEO 500-270=230 DISPLAY 125 SEO 350-125=225 SEM 500-350=150 SEO 50 SEM 375-50=325 DISPLAY 500-375=125 SEO 50 DISPLAY 350-50=300 SEM 500-350=150 Calculating the Shapley Value
  • 13.
    Calculating the ShapleyValue SEM 100 DISPLAY 270-100=170 SEO 500-270=230 SEM 100 SEO 375-100=275 DISPLAY 500-375=125 DISPLAY 125 SEM 270-125=145 SEO 500-270=230 DISPLAY 125 SEO 350-125=225 SEM 500-350=150 SEO 50 SEM 375-50=325 DISPLAY 500-375=125 SEO 50 DISPLAY 350-50=300 SEM 500-350=150 SEM’s expected marginal contribution is: DISPLAY’s expected marginal contribution is: SEO’s expected marginal contribution is: SEM 162 (0.32) DISPLAY 162 (0.32) SEO 176 (0.36) 500 orders attributed to the three channels :
  • 14.
  • 15.
    Target population Treatment1 Dead Alive Treatment 2 Dead Alive Event Event TIME What is the patient’s probability to be still alive after 20 years? Survival Analysis
  • 16.
    Visitors Channel 1 Event Conversion NO Conversion Channel 2 Conversion NO Conversion Event TIME What is the visitors’ probability to convert after 30 days or 5 visits? Survival Analysis
  • 17.
    Day1 Day2 Day5Day2 Day3 Day7 Day8 DAY 1 / VISIT 1 DAY 3 / VISIT 3 0 1 0 1 DAY 6 / VISIT 6 0 1 DAY 9 / VISIT 9 0 1 Survival Analysis
  • 18.
    VISITORS TIME STARTEND : censored observation : event (conversion) Censored observation: There is not “time to event” recorded because: •Loss of follow up  Drop out  Conversion due to a cause that is out of our interest •End of the study Survival Analysis
  • 19.
    Survival Analysis Estimatetime-to-event for a group of individuals, such as time until a visitor purchases To compare time-to-event between two or more groups, such as visitors that have clicked on a Display ad compared to visitors that have not clicked on a Display ad. To assess the relationship of co-variables to time-to-event, such as: does number of clicks, pages viewed, or time on site effect the decision to purchase?
  • 20.
    DATA DRIVEN MODELS ADJUSTED TO EACH CASE LIMITED TUTORIALS ADJUST FORMULAS TO YOUR DATA (QUITE) EASY TO SET UP LINK ADS WITH ONSITE BEHAVOR FIND COST EFFICIENT CLICK CHAINS MERGE OFFLINE AND ONLINE DATA Evaluation & Suggestions
  • 21.
    Thank you! AspaLekka a.lekka@foodpanda.com