ML for Player Retention Prediction in a Mobile Game

Client

The client is top art production and game development company in Europe. They have a proven track record in delivering hyper-casual games. The client faced the problem of predicting player retention in the new free-to-play sports game (F2P). Specifically, they had difficulty interpreting gameplay data from the first-time experience. The goal was to analyze player-behavior data and the key features influencing player churn. With all this data in place, they could reignite the player’s interest in the game and make them remain in the game longer.

The client didn’t have in-house analytics. They approached InData Labs as they were looking for a company with senior-level expertise in predictive analytics and machine learning-driven analysis. To address their business challenges, they needed intelligent predictive software.

Challenge: player retention prediction based on user-behavior data

During the planning phase, our team was challenged to develop a solution infused with efficient algorithms that analyze gameplay data and predict player churn. The solution had to be deployed into a game client.

Gameplay data is volatile, and hard to deal with. Another challenge for us was to learn how to collect this data and apply AI to predict the user’s next game move.

Solution: predictive analytics model for player churn prediction

Throughout the project, our team closely communicated with the client. We spent a lot of time in the rooms pitching ideas and discussing possible solutions. As a result, we’ve proposed adopting AI solutions for user activity analysis and churn prediction.

Our team of engineers decided on using Machine Learning and AI for proof of concept stage.

We’ve plotted out the following tasks:

• Building scoring models that determine the key player metrics within 3 months
• Predicting player churn after the first-time session, the first and second days
• Predicting the number of clicks on the ads

1. The first step was to develop high-capacity data-mining algorithms for detecting key player metrics. Since the client didn’t want to use any ready-made predictive analytics software, we had to create custom scoring models from scratch. We built the models based on a dataset of 50,000 users from the client’s F2P sports game. The data on the number of sessions per first and second days, the time between sessions, absence time, average session time were anonymized. Any personal data was neither stored nor used during the work of the project.
2. The second step was to analyze the gameplay data and predict player churn in the wild within the first-day sessions. Before that, we trained the models to analyze user behavior and determine the risk of churn. The models label the users as ‘churned’ and ‘retained’. The information on the churned users was given to the client for further analysis.
   Another task was to train the models on how to predict retention time. Having this data, the business owner can incentivize the player to remain in the game or recommend another one. By providing the company with this data, we’ve helped them understand what was really important to their audience and how they could meet them half-way.
3. The third step was about predicting the number of ad clicks. We were analyzing the overall player activity during and predicting ad clicks based on the data we had collected.

Result: improved player retention and overall gaming experience with predictive analytics

Our team has helped the client with user retention prediction in the F2P game.

Results of the collaboration at a glance:

• Player churn prediction accuracy is 92%
• Player Lifetime value prediction accuracy is 86%
• AI proof of concept development
• In-game user behavior analysis

Start Your Next Predictive Analytics Project with InData Labs