Chen Junxin, Liu Zhiqiong, Liu Jing, Li Wang, Wang Chang
Cloud Network Operation Technology Research Institute, China Telecom Research Institute, Guangzhou, 510630, China.
Sci Rep. 2025 Jul 1;15(1):22228. doi: 10.1038/s41598-025-05260-1.
Next mobile app prediction aims to recommend the apps that users will most likely to use next based on their historical usage behavior. It is critical for optimizing app preloading strategies and personalized recommendations, enhancing the user experience on mobile devices. However, it faces fundamental challenges such as interactions sparsity, rapid expansion of the app ecosystem and long-term interest neglect. Besides, user preference changes over time and frequent application updates are also ignored in existing models. To overcome the limitations of existing methods in next-app prediction, particularly in personalized feature extraction and temporal dynamics modeling, we propose a temporal-personalized next-app prediction framework, which employs multi-perspective graph representation learning with self-attention mechanisms to enhance user and app embeddings. It can effectively capture both long-term and short-term evolving user interests in app usage, enhancing dynamic temporal features of users and apps. Moreover, it can integrate global interactions into graph representation learning by multi-perspective feature aggregations. With a context-aware attention fusion mechanism applied, we effectively integrate temporal and personalized features to user and app representations. The comprehensive embeddings are obtained to next-app prediction, which significantly improve the accuracy of next app prediction. Experimental results on real datasets demonstrate that our model outperforms other baselines.
接下来移动端应用程序预测旨在根据用户的历史使用行为推荐用户接下来最有可能使用的应用程序。这对于优化应用程序预加载策略和个性化推荐、提升移动设备上的用户体验至关重要。然而，它面临着诸如交互稀疏性、应用程序生态系统快速扩张以及长期兴趣被忽视等根本性挑战。此外，现有模型还忽略了用户偏好随时间的变化以及频繁的应用程序更新。为了克服现有方法在预测下一个应用程序方面的局限性，特别是在个性化特征提取和时间动态建模方面，我们提出了一个时间个性化的下一个应用程序预测框架，该框架采用带有自注意力机制的多视角图表示学习来增强用户和应用程序的嵌入。它能够有效地捕捉用户在应用程序使用中短期和长期不断演变的兴趣，增强用户和应用程序的动态时间特征。此外，它可以通过多视角特征聚合将全局交互集成到图表示学习中。通过应用上下文感知注意力融合机制，我们有效地将时间和个性化特征集成到用户和应用程序表示中。得到的综合嵌入用于下一个应用程序预测，这显著提高了下一个应用程序预测的准确性。在真实数据集上的实验结果表明，我们的模型优于其他基线模型。
