Application usage prediction is a crucial technique for optimizing smartphone systems. This study proposes a semantic-enhanced prediction method based on large language models to improve the understanding of user intent and generalization to out-of-distribution applications. The method used a large language model to automatically generate functional descriptions of applications. These descriptions were encoded into high-dimensional semantic vectors through an embedding model and used as semantic representations of historical usage sequences. These representations were then fed into the prediction model. To better capture temporal dynamics and semantic dependencies in user behavior, a prediction model was designed by integrating locally activated attention mechanisms with a Mamba-2 module. Experimental results showed that on a real-world application usage dataset, the proposed method increased the Top-3 accuracy from 76.41% to 77.00% on in-distribution data and from 66.06% to 71.57% on out-of-distribution data, significantly improving the model's generalization to unseen applications.
针对现有方法在用户意图理解不充分与分布外泛化能力不足两方面的局限,本文提出了一种基于大语言模型(Large Language Model, LLM)语义增强的移动应用使用预测方法AppSE(Application Usage Prediction via Semantic Enhanced),通过引入语义增强机制,为预测模型提供更细粒度、通用性更强的特征表达,从而增强模型对用户行为的建模能力与对新APP的泛化能力。AppSE充分利用LLM强大的自然语言理解和生成能力,为APP生成功能描述文本,通过文本向量表示实现语义增强,在此基础上使用轻量高效预测模型实现高效端侧推理。本文主要贡献如下:
尽管上述方法在用户行为建模方面取得了显著进展,但仍主要依赖APP标识符、类别标签或人工特征,难以从语义层面反映用户行为背后的动机,限制了模型对分布外场景的泛化能力。Khaokaew等[20]提出MALPE(Mobile App Prediction Leveraging Large Language Model Embreddings)模型,首次将LLM应用于APP使用预测,将用户历史行为序列格式化为自然语言输入,利用LLM推理能力直接输出预测结果,在分布外任务中表现出较强适应性。
以T5[25](Text-to-Text Transfer Transformer,T5)模型为例,我们首先对输入的APP描述文本进行预处理,包括分词和规范化处理。处理后的文本被传递给T5模型进行前向传播,模型输出每个位置在最后一层的隐藏状态向量,其中表示第个位置的语义表示。为了消除填充部分对语义向量计算的干扰,我们引入掩蔽标识向量指示有效位置,并采用掩蔽均值池化策略计算最终的语义向量:
为提升模型对用户动态意图的建模能力,本文引入基于结构化状态空间模型(Structured State Space Models, SSMs)范式的Mamba-2[27]模块。针对APP使用预测中用户行为的时序特性,Mamba-2通过半可分矩阵的低秩分解与选择性记忆机制,增强对长期行为模式和意图迁移的感知能力。相较于传统Transformer架构,Mamba-2采用线性变换机制,将序列建模的计算复杂度从二次降为线性,显著降低了计算开销。
YangK, ZhaoX, ZouJ H, et al. ATPP: A Mobile App Prediction System Based on Deep Marked Temporal Point Processes[C]//2021 17th International Conference on Distributed Computing in Sensor Systems (DCOSS). Pafos: IEEE, 2021: 83-91. DOI: 10.1109/DCOSS52077.2021.00028 .
[6]
ChenX L, WangY, HeJ Y, et al. Cap: Context-aware App Usage Prediction with Heterogeneous Graph Embedding[J]. Proc ACM Interact Mob Wearable Ubiquitous Technol, 2019, 3(1): 1-25. DOI: 10.1145/3314391 .
KamisakaD, MuramatsuS, YokoyamaH, et al. Operation Prediction for Context-aware User Interfaces of Mobile Phones[C]//2009 Ninth Annual International Symposium on Applications and the Internet. Bellevue: IEEE, 2009: 16-22. DOI: 10.1109/SAINT.2009.12 .
[12]
LerouxP, RoobroeckK, DhoedtB, et al. Mobile Application Usage Prediction Through Context-based Learning[J]. J Ambient Intell Smart Environ, 2013, 5(2): 213-235. DOI: 10.3233/ais-130199 .
[13]
ShinC, HongJ H, DeyA K. Understanding and Prediction of Mobile Application Usage for Smart Phones[C]//Proceedings of the 2012 ACM Conference on Ubiquitous Computing. Pittsburgh: ACM, 2012: 173-182. DOI: 10.1145/2370216.2370243 .
[14]
HuangK, ZhangC H, MaX X, et al. Predicting Mobile Application Usage Using Contextual Information[C]//Proceedings of the 2012 ACM Conference on Ubiquitous Computing. Pittsburgh: ACM, 2012: 1059-1065. DOI: 10.1145/2370216.2370442 .
[15]
LiaoZ X, LiS C, PengW C, et al. On the Feature Discovery for App Usage Prediction in Smartphones[C]//2013 IEEE 13th International Conference on Data Mining. Dallas: IEEE, 2013: 1127-1132. DOI: 10.1109/ICDM.2013.130 .
[16]
ZhaoS, LuoZ L, JiangZ W, et al. AppUsage2Vec: Modeling Smartphone App Usage for Prediction[C]//2019 IEEE 35th International Conference on Data Engineering (ICDE). Macau: IEEE, 2019: 1322-1333. DOI: 10.1109/ICDE.2019.00120 .
[17]
XuS J, LiW Z, ZhangX, et al. Predicting and Recommending the Next Smartphone Apps Based on Recurrent Neural Network[J]. CCF Trans Pervasive Comput Interact, 2020, 2(4): 314-328. DOI: 10.1007/s42486-020-00045-z .
[18]
ZhangY Q, KangM Y, LiX C, et al. Optimizing Smartphone App Usage Prediction: A Click-through Rate Ranking Approach[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Barcelona: ACM, 2024: 6281-6290. DOI: 10.1145/3637528.3671567 .
[19]
WangY Z, JiangR H, LiuH C, et al. Sequence-graph Fusion Neural Network for User Mobile App Behavior Prediction[M]//Machine Learning and Knowledge Discovery in Databases: Applied Data Science and Demo Track. Cham: Springer Nature Switzerland, 2023: 105-121. DOI: 10.1007/978-3-031-43427-3_7 .
WangR, GaoZ Y, ZhangL Y, et al. Empowering Large Language Models to Edge Intelligence: A Survey of Edge Efficient LLMS and Techniques[J]. Comput Sci Rev, 2025, 57: 100755. DOI: 10.1016/j.cosrev.2025.100755 .
[22]
LiuP F, YuanW Z, FuJ L, et al. Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing[J]. ACM Comput Surv, 2023, 55(9): 1-35. DOI: 10.1145/3560815 .
[23]
ChenJ Z, FanT T, LiuX, et al. Seed 1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning[PP/OL]. (2025-04-23)[2025-08-10].
RaffelC, ShazeerN, RobertsA, et al. Exploring the Limits of Transfer Learning with a Unified Text-to-text Transformer[J]. J Mach Learn Res, 2020, 21(140): 1- 67.
[26]
ZhouG R, ZhuX Q, SongC R, et al. Deep Interest Network for Click-through Rate Prediction[C]//Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. London: ACM, 2018: 1059-1068. DOI: 10.1145/3219819.3219823 .
[27]
DaoT, GuA. Transformers Are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality[C]//Proceedings of the 41st International Conference on Machine Learning. Vienna: PMLR, 2024: 10041-10071. DOI: 10.5555/3692070.3692469 .
[28]
GuoD Y, YangD J, ZhangH W, et al. Deepseek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning[PP/OL]. (2025-01-29)[2025-08-10].
[29]
DevlinJ, ChangM W, LeeK, et al. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding[C]//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Minneapolis: ACL, 2019: 4171-4186. DOI: 10.18653/v1/N19-1423 .
[30]
RadfordA, WuJ, ChildR, et al. Language Models Are Unsupervised Multitask Learners[R]. San Francisco: OpenAI, 2019: 160025533. https://api.semanticscholar.org/CorpusID:160025533.
[31]
XUM. Text2vec: A Tool for Text to Vector[CP/OL]. (2023-09-20)[2025-08-10].
[32]
ZhangY Z, LiM X, LongD K, et al. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models[PP/OL]. (2025-06-11)[2025-08-10].