Sequential Recommendation in Online Games with Multiple Sequences, Tasks and User Levels

Si Chen, Yuqiu Qian, Hui Li, Chen LinSSTD 2021 · DOI · OpenReview

Abstract

Online gaming is growing faster than ever before, with increasing challenges of providing better user experience. Recommender systems (RS) for online games face unique challenges since they must fulfill players’ distinct desires, at different user levels, based on their action sequences of various action types. Although many sequential RS already exist, they are mainly single-sequence, single-task, and single-user-level. In this paper, we introduce a new sequential recommendation model for multiple sequences, multiple tasks, and multiple user levels (abbreviated as M3Rec) in Tencent Games platform, which can fully utilize complex data in online games. We leverage Graph Neural Network and multi-task learning to design M3Rec in order to model the complex information in the heterogeneous sequential recommendation scenario of Tencent Games. We verify the effectiveness of M3Rec on three online games of Tencent Games platform, in both offline and online evaluations. The results show that M3Rec successfully addresses the challenges of recommendation in online games, and it generates superior recommendations compared with state-of-the-art sequential recommendation approaches.

Introduction

Most existing Sequential Recommender Systems (SRS) operate under a paradigm the authors define as $S^3Rec$: Single-sequence, Single-task, and Single-user-level. These models typically analyze a single chronological stream of interactions (e.g., purchase history) to predict a single future event for an individual user.

However, modern online games, particularly sandbox titles like LEGO®Cube, present a far more complex behavioral landscape. Players do not just "interact" with items; they download maps, play maps, subscribe to categories, and engage in social interactions. These represent distinct but interlinked sequences. Furthermore, gaming is inherently social, where individual preferences are often influenced by group or guild dynamics. Traditional models that flatten these diverse signals into a single sequence or ignore social context fail to capture the nuances of player intent.

Comparison of traditional single-sequence recommendation (a) and the multi-sequence, multi-task approach of M3Rec (b) in the context of LEGO®Cube.

Methodology: The M3Rec Framework

To overcome the limitations of $S^3Rec$, we propose M3Rec, a neural architecture designed for multiple sequences, tasks, and user levels. The framework is built upon three primary pillars:

1. Gated Graph Sequence Neural Networks (GGS-NN)

Rather than treating player history as a linear list, M3Rec models interactions as a graph. This approach allows the system to capture complex, non-linear transitions between items—for instance, identifying common patterns where players switch between "Racing" and "Adventure" map categories.

2. Multi-Task Learning (MTL)

The architecture employs parallel GGS-NN units to predict multiple objectives simultaneously. While the primary task might be predicting the next map to download, auxiliary tasks predict the next map category or social interaction. These units share a common embedding layer, enabling the model to transfer knowledge across different action types.

3. Hierarchical Multi-level Modeling

To account for social dynamics, M3Rec incorporates a multi-level component. Users are clustered based on behavioral similarity, and an attention-based aggregation layer is used to derive a group-level representation. This allows the model to balance individual preferences with the collective interests of a player's social circle.

The overall architecture of M3Rec, illustrating the flow from sequential data through GGS-NN layers and the multi-task learning module.

Experimental Evaluation

The model was evaluated using three large-scale datasets from Tencent Games: LEGO®Cube, Tgame, and Ygame. These datasets encompass tens of thousands of users and a wide variety of action sequences, providing a rigorous testbed for multi-sequence modeling.

Offline Results

M3Rec consistently outperformed state-of-the-art baselines, including GRU4Rec and SR-GNN. On the LEGO®Cube dataset, M3Rec achieved a Hit Ratio (HR@10) of 0.1423, a significant improvement over SR-GNN's 0.1088. The inclusion of auxiliary tasks proved to be a "force multiplier," where social signals significantly enhanced the accuracy of content recommendations.

Online Impact

In live A/B tests on Tencent's production platforms, M3Rec demonstrated substantial gains. In LEGO®Cube, the model achieved a 100% relative improvement in key engagement metrics compared to existing production baselines. In RPG titles (Tgame and Ygame), it drove significant increases in Average Revenue Per User (ARPU) and Conversion Rates (CR).

Conclusion and Future Work

M3Rec successfully addresses the challenges of online game recommendation by leveraging the richness of multi-sequence data and social context. Future research will focus on improving model interpretability to provide actionable insights for game designers and exploring dynamic clustering methods to better track evolving player behaviors.

Summary updated

Abstract

Online gaming is growing faster than ever before, with increasing challenges of providing better user experience. Recommender systems (RS) for online games face unique challenges since they must fulfill players' distinct desires, at different user levels, based on their action sequences of various action types. Although many sequential RS already exist, they are mainly single-sequence, single-task, and single-user-level. In this paper, we introduce a new sequential recommendation model for multiple sequences, multiple tasks, and multiple user levels (abbreviated as M$^3$Rec) in Tencent Games platform, which can fully utilize complex data in online games. We leverage Graph Neural Network and multi-task learning to design M$^3$Rec in order to model the complex information in the heterogeneous sequential recommendation scenario of Tencent Games. We verify the effectiveness of M$^3$Rec on three online games of Tencent Games platform, in both offline and online evaluations. The results show that M$^3$Rec successfully addresses the challenges of recommendation in online games, and it generates superior recommendations compared with state-of-the-art sequential recommendation approaches.
BibTeX
@inproceedings{chen2021sequential,
  author = {Si Chen and Yuqiu Qian and Hui Li and Chen Lin},
  title = {{Sequential Recommendation in Online Games with Multiple Sequences, Tasks and User Levels}},
  booktitle = {SSTD},
  year = {2021},
  doi = {10.1145/3469830.3470906},
}
Figure