https://github.com/eeghor/mta

Multi-Touch Attribution
https://github.com/eeghor/mta

marketing markov-model probabilistic-models shapley

Last synced: 3 months ago
JSON representation

Multi-Touch Attribution

• Host: GitHub
• URL: https://github.com/eeghor/mta
• Owner: eeghor
• Created: 2018-10-31T07:08:10.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2022-02-26T11:33:50.000Z (over 2 years ago)
• Last Synced: 2024-05-14T00:11:30.831Z (6 months ago)
• Topics: marketing, markov-model, probabilistic-models, shapley
• Language: Python
• Homepage:
• Size: 344 KB
• Stars: 97
• Watchers: 12
• Forks: 32
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-marketing-machine-learning - mta

README

        
# mta

Multi-Touch Attribution. Find out which channels contribute most to user conversion. 

### Models

This package contains implementations the following Multi-Touch Attribution models:

* Shapley 

* Markov

* So-called Simple Probabilistic Model by Shao and Li

* Bagged Logistic Regression by Shao and Li

* Additive Hazard (Survival)

In addition, some popular heuristic “models” are included, specifically

* First Touch

* Linear

* Last Touch

* Time Decay

* Position Based

### Included Data

The package comes with the same test data set as an R package called [ChannelAttribution](https://cran.r-project.org/web/packages/ChannelAttribution/ChannelAttribution.pdf)  - there are 10,000 rows containing customer journeys across 12 channels: alpha, beta, delta, epsilon, eta, gamma, iota, kappa, lambda, mi, theta and zeta.

![data_snippet](img/data_snippet.png)

These are conversion aggregations by path. Suppose  there’s a path  (customer journey)

```

a > b > c

```

with **total_conversions** equal to 2 and **total_null** equal to 5. This means that we recorded 2 consumer journeys

```

a > b > c > (conversion)

```

and 5 customer journeys 

```

a > b > c > (null)

```

There’s an option to generate timestamp data if you want to use the Additive Hazard model (the only model that explicitly incorporates exposure times).

### References

* Nisar and Yeung (2015) - Purchase Conversions and Attribution Modeling in Online Advertising: An Empirical Investigation [pdf](https://eprints.soton.ac.uk/380534/1/GHLEFMG_FGMJHM_VJ1QM9QF.pdf)

* Shao and Li (2011)  - Data-driven Multi-touch Attribution Models [pdf](http://www0.cs.ucl.ac.uk/staff/w.zhang/rtb-papers/data-conv-att.pdf)

* Dalessandro et al (2012) - Causally Motivated Attribution for online Advertising [pdf](https://dstillery.com/wp-content/uploads/2016/07/CAUSALLY-MOTIVATED-ATTRIBUTION.pdf)

* Cano-Berlanga et al  (2017) - Attribution models and the Cooperative Game Theory [pdf](https://www.recercat.cat/bitstream/handle/2072/290758/201702.pdf?sequence=1)

* Ren et al (2018) - Learning Multi-touch Conversion Attribution

with Dual-attention Mechanisms for Online Advertising [pdf](https://arxiv.org/pdf/1808.03737.pdf)

* Zhang et al (2014)  - Multi-Touch Attribution in Online Advertising with Survival Theory [pdf](http://www0.cs.ucl.ac.uk/staff/w.zhang/rtb-papers/attr-survival.pdf)

* Geyik et al (2014) - Multi-Touch Attribution Based Budget Allocation in Online Advertising [pdf](https://arxiv.org/pdf/1502.06657.pdf)