Early churn prediction in freemium game mobile using Transformer-based architecture for tabular data

Abstract

This study introduces an application of the Feature Tokenizer Transformer architecture, a specialized adaptation of the well-known Transformer for tabular data, to predict early-stage churn in freemium mobile games. Leveraging a comprehensive dataset of over 268,370 users, we focus on the meticulous preprocessing of minimal feature sets that are available immediately after post-tutorial phase of user’s onboarding. These features, easily collectible through any Application Data Monitoring system, include device information, user demographics, user acquisition data, user interactions during tutorial steps, and initial session behavior. We propose an novel approach for imputing missing numerical data values by developing Gradient Boosted Regression models for each individual column within the dataset. This approach is designed to understand the complexity relationship between values in columns, allowing to predict the null values accurately rather than simply filling them with mean or zero values. Despite the inherent challenges associated with the FT-Transformer, such as extensive training times, hyperparameter tuning complexities, and a substantial data requirement, our methodology achieved an Area Under ROC Curve of 94.9%, an accuracy of 86.79% and recall of 88.75%, demonstrates significant potential in enhancing churn prediction accuracy. While eXtreme Gradient Boosting currently surpasses the FT-Transformer in performance, our approach still outperforms existing methods in early churn detection for freemium mobile games. To the best of our knowledge, this is the first study to explore the potential of the FT-Transformer in freemium mobile game early churn detection, achieving high performance with a generalized dataset adaptable to various applications. We also introduce a machine learning-based imputation strategy for missing data, addressing a common problem in early churn prediction.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 5 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.