Abstract
Mobile gaming has emerged as a promising market with billion-dollar revenues. A variety of mobile game platforms and services have been developed around the world. One critical challenge for these platforms and services is to understand user churn behavior in mobile games. Accurate churn prediction will benefit many stakeholders such as game developers, advertisers, and platform operators. In this paper, we present the first large-scale churn prediction solution for mobile games. In view of the common limitations of the state-of-the-art methods built upon traditional machine learning models, we devise a novel semi-supervised and inductive embedding model that jointly learns the prediction function and the embedding function for user-app relationships. We model these two functions by deep neural networks with a unique edge embedding technique that is able to capture both contextual information and relationship dynamics. We also design a novel attributed random walk technique that takes into consideration both topological adjacency and attribute similarities. To evaluate the performance of our solution, we collect real-world data from the Samsung Game Launcher platform that includes tens of thousands of games and hundreds of millions of user-app interactions. The experimental results with this data demonstrate the superiority of our proposed model against existing state-of-the-art methods.
Introduction
Predicting user churn—when a user stops playing a game—is a vital task for mobile gaming platforms to optimize retention and marketing. Traditional models like Random Forests or Logistic Regression often struggle to scale to platforms hosting tens of thousands of apps and millions of users. Moreover, existing graph embedding techniques are frequently transductive, meaning they cannot easily handle new users or games that appear after training. This paper addresses these challenges by proposing a novel semi-supervised and inductive graph embedding model.
Problem Formulation
The interaction between users and games is modeled as an attributed bipartite graph $G^{(t)}$, where nodes represent players and games, and edges represent play records at time $t$. Each node is associated with a set of attributes, such as device type for players or genre for games. This edge-centric approach captures the specific relationship between a user and a game, which is more predictive of churn than node-centric features alone.
Proposed Methodology
The core of the solution is a Deep Neural Network (DNN) architecture that learns an inductive edge embedding function. Unlike transductive methods that learn fixed embeddings for specific IDs, this model learns a function based on node attributes, allowing it to generate embeddings for brand-new users or games on the fly.
The model employs a semi-supervised joint training approach, combining two primary tasks:
- Unsupervised Context Inference: This component learns graph context by predicting which edges are likely to appear near each other. It uses a novel Attributed Random Walk that considers both topological connections and attribute similarities, allowing the model to learn relationships between similar users even if they haven't played the same games.
- Supervised Churn Prediction: This component directly predicts the probability of churn for a given user-game relationship.
Additionally, the model incorporates Temporal Loss to account for the evolving nature of user behavior and "right censoring," a statistical challenge where recent data lacks definitive labels because the churn window has not yet passed.
Experimental Evaluation
The model was evaluated on two massive datasets from the Samsung Game Launcher platform, representing some of the largest-scale tests for churn prediction in the literature:
- USA Dataset: 15,000 users, 19,705 games, and ~76.5 million play records.
- Korea Dataset: 25,000 users, 18,470 games, and ~106.5 million play records.
The proposed semi-supervised (SS) model significantly outperformed traditional machine learning baselines. In the USA dataset, it achieved an AUC of 0.82 and a Recall of 0.78, compared to the strongest baseline (Random Forest) which achieved an AUC of 0.75 and a Recall of only 0.31. The dramatic improvement in Recall—over 2.5x—is particularly valuable for business applications, as it allows platforms to identify and target a much larger proportion of actual churners.
Conclusion
This work presents a scalable, inductive solution for churn prediction in large-scale mobile games. By focusing on edge-centric embeddings and combining supervised and unsupervised learning, the model captures complex user-game dynamics effectively. While limited by the lack of granular in-game data due to privacy, the framework is highly generalized and can be applied to other bipartite churn problems, such as subscription cancellations or social network group unsubscriptions.
Summary updated