1
$\begingroup$

As you might already know there is a concept of retention. Let's say I have created a game and today hundred people have downloaded my game. Let's say tomorrow 47 out of yesterday's hundred people are returning to the game, i.e. they open the game and play it. This means that day one (D1) retention is 47%.

Now let's say I want to increase my D1 retention from 47 to 60. Let's say I can mine a classification formula or a classification tree that when I give a D0 user data, that algorith can say if that player will return on D1.

How can I potentially use that classification model to increase D1? Or let's put it another way, what is the proper way using data science to understand what to do to increase the retention?

$\endgroup$
6
  • $\begingroup$ I think that this field of research is called 'churn prediction'. An example application is to predict which customers will not return for D1 and tell the business to do something about it e.g. make a promotional offer or targeted marketing. $\endgroup$
    – Enk9456
    Commented Jan 22, 2023 at 16:09
  • $\begingroup$ @Enk9456 does it mean that I can only predict that this user will churn, but I do not have any reasonable way to stop the train based on the prediction model. In other words is there any useful information in my model telling me how exactly to prevent that user from churn. $\endgroup$
    – Narek
    Commented Jan 22, 2023 at 16:15
  • $\begingroup$ You must provide additional information about the features you use to model your problem; otherwise, it makes no sense. Please edit your question, and I may try to formalize an answer. $\endgroup$
    – Eduard
    Commented Jan 22, 2023 at 16:41
  • $\begingroup$ @0xedu thanks for commenting. This is a general question. I'm new to data science. And want to understand if there is a methodology not only to predict a churn but also based on that prediction model try to stop the churn. Is there any algorithm, any methodology, any approach to accomplish this kind of tasks? $\endgroup$
    – Narek
    Commented Jan 22, 2023 at 16:50
  • $\begingroup$ The literature is flourishing with approaches for churn prediction and, thus, learning algorithms. First, however, you must understand your data and then hope to apply some machine learning algorithm. Recall that "garbage in, garbage out" still reigns. If you want references on churn prediction, I may try to point you to some. $\endgroup$
    – Eduard
    Commented Jan 22, 2023 at 16:55

1 Answer 1

1
$\begingroup$

You can get inspiration from reading the following papers: 1, 2, 3, 4, 5

In any case, I recommend you investigate more on your data. In particular, you should be able to tell a story before employing any machine learning algorithm. For example, you should ask yourself the following questions:

  • Can I have natural labels (e.g., labels that your users provide by clicking here instead of there) in my dataset? For example, in recommendation systems, you can easily have natural labels (e.g., the user chooses character A over character B to stay in your game context).
  • What features play a crucial role in my business domain (e.g., gender)? It should be noted that business metrics, such as retention (as you have mentioned), are different from academic metrics, such as accuracy and F1 score, to name a few.
  • What is the distribution of examples if I segment my data, for example, based on gender, years, and location? Of course, knowing your users is always a good thing.
  • Are the features correlated? This is essential in a high dimensional space where feature interaction may eventually fool your machine learning algorithm.
  • Can I group (near)duplicated examples? This is also crucial because you must prepare your dataset by putting together such cases in the same set.
  • How can I avoid data leakage when preparing the data? Data leakage is a worrying phenomenon in data science and machine learning because it simply says that you should avoid poisoning your predictor variables with the target's information. Moreover, this is a challenging issue when dealing with time series data.
  • ...

There is no general rule to obtain the ideal data for training your models. Nevertheless, as I understand your problem, you should start with your business requirements and lead the way to exciting inquiries.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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