Large-scale User Game Lifecycle Representation Learning

TL;DR

This paper tackles the sparsity and long-tail imbalance in game advertising and in-game item recommendation by introducing the User Game Lifecycle (UGL), a self-supervised representation learning framework that enriches user actions by aggregating signals from multiple environments. It adds action augmentation through Aggregation and Negative Feedback, and balances training through Inverse Probability Masking (IPM), enabling effective learning from scarce and skewed game interactions. The model consists of a sequence embedder and a Transformer-based sequence encoder, with four pooling strategies to form a unified UGL representation used by downstream task models; extensive offline and online experiments on Tencent-scale data show consistent improvements in AUC for advertising (1.53–2.08%) and CVR/ARPU for recommendations (CVR +21.67%, ARPU +0.82%). The work demonstrates scalable, domain-specific representation learning that improves both targeting accuracy and monetization signals, offering practical benefits for large game ecosystems with limited item catalogs and highly imbalanced user behaviors.

Abstract

The rapid expansion of video game production necessitates the development of effective advertising and recommendation systems for online game platforms. Recommending and advertising games to users hinges on capturing their interest in games. However, existing representation learning methods crafted for handling billions of items in recommendation systems are unsuitable for game advertising and recommendation. This is primarily due to game sparsity, where the mere hundreds of games fall short for large-scale user representation learning, and game imbalance, where user behaviors are overwhelmingly dominated by a handful of popular games. To address the sparsity issue, we introduce the User Game Lifecycle (UGL), designed to enrich user behaviors in games. Additionally, we propose two innovative strategies aimed at manipulating user behaviors to more effectively extract both short and long-term interests. To tackle the game imbalance challenge, we present an Inverse Probability Masking strategy for UGL representation learning. The offline and online experimental results demonstrate that the UGL representations significantly enhance model by achieving a 1.83% AUC offline increase on average and a 21.67% CVR online increase on average for game advertising and a 0.5% AUC offline increase and a 0.82% ARPU online increase for in-game item recommendation.

Figures (7)

• Figure 1: An example of User Game Lifecycle with Aggregation and Negative Feedback strategies for user interest induction. $w_{(t,g)}$ is the threshold of the pair of the action type $t$ and the game $g$, and $w_s$ is the threshold for silent actions.
• Figure 2: The distribution of action types across different games, where $(t,g)$ represents the action with type $t$ in the context of game $g$. Notably, actions associated with popular games (on the left side) exhibit a significant prevalence within the sequence data, while actions related to cold games (on the right side) are relatively sparse.
• Figure 3: Illustration of Inverse Probability Masking (IPM), a model structure, and the application of UGL rep. in a task-specific model. In the model structure, $t^g_i$ is masked as [M] during the training stage, and $t^g_i$ remains unmasked during the inference stage. Notably, in the training stage, IPM is applied to optimize UGL representation learning. The UGL rep. obtained in the inference stage is fed to a task-specific model for a downstream advertising/recommendation task.
• Figure 4: Illustration of the Sequence Embedder. The input embedding is computed as the sum of embeddings for type-game, start and end dates, and frequency.
• Figure 5: Training losses (solid) and accuracy (dashed) of UGL representation with IPM and with vanilla methods.