[Paper Review] A Semi-Supervised and Inductive Embedding Model for Churn Prediction of Large-Scale Mobile Games

Xi Liu, Muhe Xie+5 more|arXiv (Cornell University)|Aug 20, 2018

Customer churn and segmentation4 citations

TL;DR

This paper proposes a novel semi-supervised, inductive deep learning model that jointly learns user-game interaction embeddings and churn prediction functions using attributed edge embeddings and temporal loss. Evaluated on a real-world dataset of 40,000 games and 40,000 users from Samsung Game Launcher, the model outperforms state-of-the-art methods in large-scale churn prediction by capturing both contextual features and dynamic relationship changes through a novel attributed random walk and edge embedding framework.

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.

Motivation & Objective

• To address the critical challenge of churn prediction in large-scale mobile gaming platforms with billions of user-app interactions.
• To overcome limitations of traditional machine learning models, including poor scalability, reliance on handcrafted features, and inability to handle new users and games.
• To develop an inductive, semi-supervised model that jointly learns prediction and embedding functions for user-game relationships.
• To incorporate rich contextual attributes (e.g., WiFi, screen brightness) and dynamic interaction patterns into the embedding process.
• To enable real-time, scalable churn prediction suitable for integration into commercial mobile game platforms.

Proposed method

• The model represents user-game interactions as an attributed bipartite graph, where nodes are users and games, and edges encode interaction events with attributes.
• A novel attributed random walk technique is designed to sample edge contexts by considering both topological adjacency and attribute similarity between interacting entities.
• Edge embeddings are learned via a deep neural network that jointly optimizes for churn prediction (supervised) and context recovery (unsupervised).
• Temporal loss is introduced to enforce consistency of edge embeddings across consecutive graph snapshots, capturing relationship dynamics over time.
• The model is inductive, enabling inference on previously unseen users, games, and user-game edges without retraining.
• The framework supports end-to-end, automatic feature learning without manual feature engineering, leveraging both structural and attribute-rich data.

Experimental results

Research questions

• RQ1 Can a semi-supervised, inductive edge embedding model effectively predict churn in large-scale mobile games with hundreds of millions of user-app interactions?
• RQ2 How does incorporating both topological structure and attribute similarity in random walk sampling improve predictive performance?
• RQ3 To what extent does modeling temporal dynamics in edge embeddings enhance churn prediction accuracy compared to static embeddings?
• RQ4 Can the proposed model generalize beyond mobile games to other churn prediction scenarios involving user-object relationships?
• RQ5 How does the joint optimization of prediction and embedding functions compare to decoupled or purely unsupervised approaches?

Key findings

• The proposed model achieves superior performance on a real-world dataset of 40,000 games and 40,000 users collected from the Samsung Game Launcher platform.
• The model significantly outperforms existing state-of-the-art methods in churn prediction accuracy, particularly in handling large-scale, dynamic user-game interactions.
• The inclusion of temporal loss leads to more stable and accurate edge embeddings, improving prediction reliability over time.
• The attributed random walk technique effectively captures richer context by combining topological and attribute-based similarities, enhancing representation quality.
• The semi-supervised training scheme improves generalization, especially in low-supervision regimes, by leveraging unlabeled data for context recovery.
• The model is fully inductive, enabling deployment in production systems where new users and games are continuously introduced.