How to Use Machine Learning Models in Mobile App Personalization

Gone are the days when machine learning was a rare technology only available to digital giants. Today, you need big data to build a successful business system in almost any industry.

Mobile app development is no exception. App owners using a data-driven approach to decision-making are increasingly convinced of its efficiency, which results in real financial gains.

Composite data analysis using ML models can help improve key app metrics, e.g. increase life time value (LTV), retention and average revenue per user (ARPU), decrease churn, etc.

One of the most practical and popular ways to leverage machine learning in a mobile app is through personalization, which is precisely what we will be discussing in this article.

We will talk about different approaches to personalization based on ML algorithms and help you select the one best suited for your business.

Why does app personalization matter?

At the heart of personalization is the simple idea of providing users with content, experience, and features that are customized to their specific needs at a particular point in time. The customized offer may include anything – a product needed “here and now”, suitable subscription terms, or, for example, a personal gift.

Personalization has been widely used in recent years, playing a major role in any digital system. It allows developers to connect with the audience and adapt to its requirements as they have never done before.

A well-personalized mobile app automatically collects user information, adjusting to their preferences without any additional action on their part.

Where to get the data from?

Using ML algorithms, including personalization models, always starts with data collection. There are many ways of learning more about your customers. Some personal information is generally provided during the sign-up process while other data may be obtained from public sources, such as social media.

Apart from that, mobile users routinely perform various actions, e.g. launching the app, viewing content, liking and sharing posts, making purchases and taking out subscriptions. These actions also characterize users and the way they interact with the product.

All data that can be collected with the app and used in personalization models fall into categories corresponding to general and user-specific factors.

User-specific (personal) factors:

General factors (affecting the entire audience):

It is essential for users to know that all data are collected anonymously. A model does not identify users.

Personalization models

We will look at several personalization models based on machine learning. For the best results, the model should be consistent with your end goal and product type.

For example, if your app has multiple subscription plans, a dynamic pricing model would be appropriate. Ranking models would be suitable for e-commerce apps dealing with a large number of products. In utilities or banking apps, the user experience can be greatly improved by predicting behavior.

The type and amount of information are also important. Different algorithms are trained on different data. The more user factors you collect, the more personalization models you can implement in your product.

Now, we are going to discuss the data required to train various models.

Dynamic pricing model

Monetization is the top priority for almost all mobile app owners and developers. A business needs to generate revenue that not only covers its costs, but also allows it to make a profit. In-app purchases remain one of the most popular sources of revenue for mobile apps, along with subscriptions.

Successful monetization through in-app purchases or subscriptions requires not only careful selection of goods and services, but also adequate pricing. You need to be aware of the following two aspects:

With an audience large enough for A/B testing, we can easily do what was almost impossible just a few years ago – personalize the prices using dynamic pricing.

In-app prices should not be rigidly fixed, fluctuating within a certain range instead. This scenario implies that the final price is set automatically for each user.

For example, the dynamic pricing system in Top War: Battle Game will set the price for the Platinum Commander item at the highest level within the RUB 799–1,990 range that the user might be willing to pay.

A dynamic pricing model may prove more efficient than a fixed price based on the sales history, competitor analysis, or a series of A/B tests, because ML models set prices in real time taking into account:

Churn prediction model

User churn is inevitable in any app’s life cycle. Not all of your customers will stay with you forever. The good news is that you can and should regulate your app’s Churn Rate.

To reduce churn, you can use a special ML model to identify users most likely to quit, and then try to retain them with customized offers.

Researchers from Samsung Research America and three US universities conducted a study to create an ML system for predicting churn in mobile games, a market worth billions of dollars.

Using the play history, game profiles (including features like genre, developer, number of downloads, rating values, and number of ratings), and user information (i.e. device model and region) helped achieve exceptional churn prediction quality. The model is applicable not only to mobile games and platforms, but to other segments as well.

User behavior prediction model

Personalization based on predicting user behavior can also be very effective. It relies on models that analyze behavior patterns. Historical data on how a user has been interacting with your app helps understand their character traits and peculiarities, which can be used to model their next steps.

Such models can “guess” the action a user is about to take at any given moment – the next purchase in the store, next money transfer through the banking app, including the amount and recipient, or switching to another playing mode while gaming.

The figure shows that at any given time user ti is most likely to perform a certain action. The app can send them an appropriate notification, or update the quick access menu accordingly.

A model is trained on user, device, and in-app behavior data.

Ranking model

Ranking models are also widely used for personalization, particularly by streaming services, social media and search engines.

The point is to show people the content they are most likely to appreciate or find useful. Ranking is most critical to e-commerce apps, where it directly affects customer conversion.

Imagine you are looking for a certain smartphone at a marketplace with tens of thousands of items. You would expect to see it among the first search results and would be disappointed if all top spots are taken by phone covers.

Of course, small stores can simply rank products by conversion. However, when there are too many products and factors influencing the purchase, using an automated model is what makes for the most effective and profitable ranking.

MyTracker Personalize

MyTracker offers app owners and developers a wide range of ML-based personalization approaches. We currently use:

MyTracker machine learning algorithms will help you personalize all your offers and in-app purchases, subscription plans, and product card rankings.

MyTracker Personalize automatically segments the audience to select the most appropriate offer for each user without the help of a game designer, product manager or manual A/B testing.

Its efficiency was confirmed by a three-month experiment at using MyTracker Personalize models in Hustle Castle, a mobile castle simulator. The personalized offers helped increase ARPU by 23% within the tested group.

Summing up

Personalization can significantly improve the key metrics of almost any product, with ML models helping to implement it in the app.

In recent years, ML models have become much simpler and no longer require major R&D. Just choose a suitable personalization type for your app and use MyTracker Personalize.

Tags: prediction monetization Personalize
  • Contents