User Representation via Cross Multi-source Behavior Pre-training for Mobile Games
A research paper titled 'User Representation via Cross Multi-source Behavior Pre-training for Mobile Games' was published on arXiv (cs.AI) on 2026-09-01. The paper proposes CM-PTM, a Cross Multi-source Behavior Pre-Training Model for mobile game user representation learning on device-level behavioral logs. CM-PTM uses hierarchical cascaded mask-then-predict proxy tasks to infer the source of the next behavior and refine predictions at the app-action level. Experiments on large-scale real-world mobile datasets demonstrate that CM-PTM effectively captures users' endogenous interests.
The paper introduces CM-PTM, a pre-training model designed to learn user representations from cross-source, multi-granular mobile device behavior logs, specifically targeting mobile games. Unlike prior work that focuses on single-app or app-level behaviors, CM-PTM models device-level interactions across heterogeneous behavior sources and hierarchical action structures. The model employs hierarchical cascaded mask-then-predict tasks: first predicting the source of the next behavior, then refining predictions at the app-action level. This unified pre-training paradigm aims to alleviate data sparsity in downstream personalization tasks. Experimental results on large-scale real-world mobile datasets show that CM-PTM effectively captures users' endogenous interests.
CM-PTM's hierarchical cascaded mask-then-predict proxy tasks enable unified modeling of cross-source dependencies and fine-grained behavioral dynamics. The two-stage prediction (source then app-action) likely forces the model to learn both coarse-grained context switching and fine-grained action sequences, potentially improving representation quality for sparse downstream tasks. The use of device-level logs suggests the model can leverage signals across multiple apps, which is a departure from app-centric pre-training.
This research addresses a gap in mobile personalization: existing pre-training methods often ignore cross-app and device-level behavior. For mobile game companies and ad platforms, better user representations from device-level logs could improve recommendation, churn prediction, and ad targeting. The approach may be applicable beyond games to any mobile app ecosystem where user behavior spans multiple apps.
For mobile game developers and publishers, improved user representations can lead to better personalization, increased user engagement, and higher monetization. For ad platforms, cross-source behavior modeling could enhance ad targeting and user segmentation. The pre-training approach may reduce the need for large amounts of labeled data in downstream tasks, lowering data collection costs.
Next observable signals include: (1) release of code or pre-trained models by the authors; (2) follow-up papers applying CM-PTM to other domains (e.g., e-commerce, social media); (3) industry adoption by mobile game publishers or ad networks; (4) benchmark comparisons against other cross-app pre-training methods. The paper's acceptance at a major conference would also be a signal of validation.