Collaborative-Enhanced Prediction of Spending on Newly Downloaded Mobile Games under Consumption Uncertainty

TL;DR

This work targets predicting spending on newly downloaded mobile games under uncertain user behavior. It introduces a robust framework that standardizes spending labels to stabilize training and evaluation, and a collaborative-enhanced model that leverages user download history without using user IDs to protect privacy, allowing online training. Empirical results show a 17.11% improvement on offline data and a 50.65% revenue uplift in an online A/B test over production models, underscoring both predictive accuracy and real-world impact. The contributions advance stable model deployment under noisy spending patterns and demonstrate practical paths for privacy-preserving collaboration signals in mobile-game recommendation and revenue optimization.

Abstract

With the surge in mobile gaming, accurately predicting user spending on newly downloaded games has become paramount for maximizing revenue. However, the inherently unpredictable nature of user behavior poses significant challenges in this endeavor. To address this, we propose a robust model training and evaluation framework aimed at standardizing spending data to mitigate label variance and extremes, ensuring stability in the modeling process. Within this framework, we introduce a collaborative-enhanced model designed to predict user game spending without relying on user IDs, thus ensuring user privacy and enabling seamless online training. Our model adopts a unique approach by separately representing user preferences and game features before merging them as input to the spending prediction module. Through rigorous experimentation, our approach demonstrates notable improvements over production models, achieving a remarkable \textbf{17.11}\% enhancement on offline data and an impressive \textbf{50.65}\% boost in an online A/B test. In summary, our contributions underscore the importance of stable model training frameworks and the efficacy of collaborative-enhanced models in predicting user spending behavior in mobile gaming.

Figures (6)

• Figure 1: This illustration showcases our collected data. It includes observed spending behavior by users leading up to an observation deadline. Users' spending habits are tracked from the day they download the game, with a 30-day observation period to gauge their spending during the subsequent 30 days. Different colored circles represent different games, with larger circles indicating higher spending. Using this observed data, we aim to predict users' future spending within the next T days for newly downloaded games.
• Figure 2: Repeated three production experiments, labeled Exp. 1, Exp. 2, and Exp. 3, were conducted using a private dataset to evaluate the spending money prediction capacity. RMSE-based loss is used as objective loss, and R2 Score Wikipedia_2023_r2score is used as model effectiveness evaluation. The worst-performing model is marked with the baseline. The numbers above each bar indicate improvements compared to the baseline. For example, the Exp. 1 bar, represents a 2.15% improvement in training loss and a 7.64% improvement in the R2 Score over the worst-performing experiment.
• Figure 3: The Comprehensive Architecture of Our Proposed Framework and Model. The modules colored other than gray represent the new modules we have introduced.
• Figure 4: Training and Evaluation Process Curve
• Figure 5: Performance of our proposed model under different label standardization types. BS denotes both-sided, GS denotes game-sided, US denotes user-sided, Log denotes logarithmic, and OV denotes origin values.