2
$\begingroup$

I would like to develop an algorithm for Multi-Touch Time-Decay Attribution Model for an e-mail campaign. I have the research paper ready and I would convert the paper into R-script. I am struggling to find the data for the purpose.

Can anyone please give me some pointers on where to find such kind of data or what kind of features are considered for such model, so that I can try to simulate the data?

Thank you in advance.

$\endgroup$
1
$\begingroup$

We have recently completed one such exercise using Markov chains. The data that we used for this project was as follows: For a selected set of users, for a 6 month time period their complete campaign history. Format as follows: Userid, campaign-id, campaign-date, response tag, response value

The response tag is a binary tag indicating responded or no.

Our markov chain approach did not really care about the time of campaign, it only used the order and getting null state vs success.

What approach are you using in your model? I'd be happy to compare notes. While I'd not be able to share the data, we could run your code on our data.

We had real campaign history data for a client, hence didn't have to simulate it.

$\endgroup$
  • $\begingroup$ Hi Tak, thank you for the reply. My model is on my Github. The link is github.com/dalalkrish/Attribution-Model. Your suggestions would be highly appreciated. I have read about Markov chains model on R-bloggers I guess there is package in R that serves the purpose. However, the model that I have used is a probabilistic model and it is also dependent only on the response value. $\endgroup$ – Krish Aug 17 '16 at 6:29
1
$\begingroup$

It was not a R specific analysis or super sophisticated, but we did use our data to analyze email performance re: lifetime value calculation and survival rate.

We took all the events (email open, clicks, unsubscribes), by a particular user and normalized to the time they first signed up. We were able to build a simple modal to predict the expected lifetime for each group of users. We could have expanded it to use it for churn prediction analysis, but at the time, we were shifting away from email marketing to the sexier social marketing so we did not continue to improve the model.

Some features that we've found useful:

  1. The sequence of the email (e.g., 1st, 2nd, 3rd)
  2. Action taken (or not taken) in the preceding email
  3. How long they have been in the database (1 day, 3 month, 2 years, etc.) This one is closely related to 1 though

etc.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.