Problem
A product team tracks many user events (opened a notification, read the blog, created a login) across several channels. It wants to know which of them lead to the target action and which users are close to it. Single-touch rules give all the credit to one event, and the linear rule over-credits small channels. At trivago the target action was a hotelier starting a new campaign in Business Studio.
Example from the article: a path "blog checked > email opened > subscription campaign started" has three events, and each attribution model splits the credit for the last one differently.
Solution in brief
Pick the attribution model by purchase cycle length, channel mix and ad spend. For a data-driven version, count converting and non-converting paths per channel and per channel pair, combine them with a second-order Markov estimate, and store the result in a lookup table. A user's score is then looked up by the path taken so far. The output is a contribution table, not a causal effect.
Questions this case answers
- How to attribute a conversion to the touchpoints on a user path when the standard single-touch rules over- or under-credit channels
- How to turn historical event paths into a per-user conversion propensity score for audience selection
- How to split credit for a conversion among the channels and events on a user's path?
- How to choose between last-touch, first-touch, linear and data-driven attribution?
Tools and methods
- Last, first and linear touch attribution — baseline rules for crediting a conversion.
- Second-order Markov chain attribution — data-driven channel contribution from single-channel and pairwise conversion probabilities.
- User conversion score — lookup of a user's path in the contribution table, scaled 0 to 1.
- pandas, NumPy — counting and the contribution calculation.
Google or a one-time email code. No password.