Blog

Q&A Session about Machine Learning

After our interview with the Machine Learning specialist, we received a bunch of questions from our customers. As this topic is apparently of great interest, we have decided to organize several Q&A sessions. 


Let's start with the first questions.



- We want to apply ML to our marketing campaigns. Can we use loyalty points to increase email open or click rates?

- You can definitely apply ML to your marketing campaigns. Of course, your issue can be solved just with statistics & A/B testing. No ML will be needed here. But more complex conditions require more scientific methods. Automatic models can increase action rates, if the dataset is clean and big enough. In the current case it is enough to conduct one experiment on the part of the target audience. You just need to do the following steps:
  • assign loyalty points to the first group of users
  • assign no points to the second group
  • calculate CTRs
  • compare data using statistical tests
If you want you can conduct multiple experiments. E.g. assign 500 points to one group, 1000 points to another group and compare them to monitor group with no points at all. The experiment is a bit more complex, but the core is the same. Using appropriate statistical criterias you can calculate the difference and significance between group CTRs. Amount of involved users allows to receive statistically significant results that can be calculated beforehand. 



- How many loyalty points should we offer a specific mailing list subscriber to get the desired engagement?

- It is a bit more classic query for ML compared with the previous one. To get a business value we should solve the following issues:
  1. Do we have enough data about users? E.g. your audience can mostly be new users with small retention rates. So we won't have any data about the difference between users. And, vice versa, your users can have a long history of using your service so ML can extract some deep knowledge from their behavior.
  2. You should be ready to conduct multiple experiments after ML launch. To train ML model, we need to know the proven data about users. But for this purpose, we can’t gather real target data. I.e. we can’t assign the same user in the same condition 100 points, then 200, 500 or 700 and see that 700 points is enough for him to act while 500 points is not enough. Thus the model will be designed and trained on some proxy metrics of "how profitable would be the distribution of X points among the users". It provides some additional uncertainty into the model that cannot be solved outside. So the result of the model has to be checked via experiments.
  3. Usually one of the main tasks is to decide what we can do with the user to push him to make an action. Or maybe we should just wait and do nothing because the user is not ready to buy. It also can be partially solved by ML. But the certain development heavily depends on the actions that model can operate with as well as on users' data that is already gathered.

Contact form

Let's discuss our cooperation
Please fill in the form below and we will contact you to clarify the details