Cross-channel Multi-device Conversion Attribution

Cross-channel Multi-device Conversion Attribution
Qingchen Wang, Marc Salomon, Ger Koole, Maarten Soomer, Kevin Pak
Are the models any good?
Conversion Attribution
Advertising campaigns are often launched across multiple channels.
Billboard, TV, radio, out of home, print, direct mailing.
Web search, online display, social media (Facebook, Twitter), video (YouTube), mobile
(browser and apps), and direct emailing.
What is the incremental effect that an advertising activity (touch point) has
on the likelihood of converting?
β€’
Credit is awarded to the advertising channels/vendors proportionally to their incremental
effects.
Attribution Results with Different Algorithms
19000
True Expected
17000
Number of Conversions Attributed
β€’
β€’
Last Touch
15000
All Touch
13000
Simple Prob.
11000
Causal CA
9000
Markov Chain
7000
Current CA
5000
Current CA-Alt
3000
Logistic Reg.
1000
Linear Reg.
-1000
TP-1%
Mean Abs
Error
Source: http://www.mobyaffiliates.com/blog/appsflyer-launches-multi-touch-attribution-analytics-to-track-mobile-app-installs/
TP-0%
Last Touch
All Touch
Simple
Probilistic
Causal CA
Markov
Chain
Current
CA
Current
CA-Alt
24320.47
24314.79
28632.73
42261.41
33987.68
14071.27
4243.42
Logistic
Linear
Regression Regression
8068.55
835.13
Effect of fragmented data
1. Last touch attribution (LTA)
Attribution Results with 20% Broken Paths
Source: http://www.clickz.com/clickz/column/2282207/embracingthe-reality-of-multitouch-attribution
4. Time-decay attribution
5. Bathtub attribution
Attribute conversion to first and last TPs
Number of Conversions Attributed
19000
2. First touch attribution
Attribute equal value to all TPs
TP-5%
Source: Shao and Li (2011)
CA in the industry (rule-based)
3. Linear attribution (all touch)
TP-2.5%
True Expected
14000
Last Touch
Simple Prob.
Causal CA
Markov Chain
9000
Current CA
Current CA-Alt
Logistic Reg.
Linear Reg.
4000
6. Engagement mapping
Attribute conversion based on userspecified weights
-1000
Source: http://www.clickz.com/clickz/column/2282207/embracing-the-reality-of-multitouch-attribution
TP-1%
Mean Abs
Error
TP-1%
Simple
Prob
Causal CA
Markov
Chain
Current
CA
Current
CA-Alt
29570.26
22501.12
26867.18
27053.08
31224.05
17921.30
11052.31
Markov chain (Anderl et al 2013)
Treat each TP as states in a Markov chain with conversion as the absorbing state.
Consumers move from one state to another down the conversion path
Hidden Markov models (Abhishek et al 2012)
Consumers have latent intrinsic states (disinterested
based on which state he/she is in
interested) and reach TPs
Vector-Autoregressive model (Wiesel et al 2011)
TP-0%
Logistic
Linear
Regression Regression
15043.63
10418.12
Attribution Results with Different Algorithms
0.1300
Number of Conversions Attributed
Treat CA as a survival analysis problem with conversion as death and estimates the
effect of each TP on the rate of death
TP-0%
Effect of targeting
Computes the empirical 𝑃 π‘π‘œπ‘›π‘£π‘’π‘Ÿπ‘ π‘–π‘œπ‘› 𝑇𝑃𝑖 ) for all 𝑖 and then get the relative
contribution of each TP
Survival analysis (Manchanda et al 2006, Papelnjak 2010)
TP-5%
All Touch
Simple probabilistic model (Shao and Li 2011)
Treat CA as a binary prediction problem and use the estimated coefficients to
derive the marginal contribution of each TP
TP-5%
Last Touch
CA in papers (data-driven)
Logistic regression (Li and Kannan 2014, Nottorf 2014)
TP-2.5% TP-2.5%
0.1100
True Expected
Last Touch
0.0900
All Touch
Simple Prob.
0.0700
Causal CA
Markov Chain
Current CA
0.0500
Current CA-Alt
Logistic Reg.
0.0300
Linear Reg.
0.0100
-0.0100
TP-0%
Treat CA as a time series problem with advertisement channel spend and sales
dependent on spend and sales in previous time step. Uses aggregate-level data
Mean Abs
Error
TP-3%
Retarget-0% Retarget-3% Search-0% Search-3%
Last Touch
All Touch
Simple
Prob
Causal CA
Markov
Chain
Current
CA
Current
CA-Alt
0.0779
0.0628
0.5605
0.0998
0.4833
0.0456
0.0281
Logistic
Linear
Regression Regression
0.0447
0.0039