GPTCN: Gated Parallel Transformer Convolutional Networks for Downstream-Task User Representation Learning on App Usage

• Published in: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 5175 - 5179
• Main Authors: Sun, Yingjie, Zeng, Fanrui, Xiao, Jiamin, Deng, Yuxiao, Ding, Yifan, Li, Yizhou
• Format: Conference Proceeding
• Language: English
• Published: IEEE 14.04.2024
• Subjects: fine-tuning; Manuals; Mobile applications; mobile data modeling; Representation learning; Signal processing; transformer; Transformers; User experience; user representation learning; Vectors
• ISSN: 2379-190X
• DOI: 10.1109/ICASSP48485.2024.10446256

With the development of mobile applications into a part of modern life, the user usage behavior data of mobile applications can well reflect the attribute characteristics of users. For many downstream applications, including advertising, recommendations provide effective support. To provide users with customized services and optimize user experience, industry and scholars have been exploring feasible solutions. However, automatic user modeling based on mobile app usage faces unique challenges, including (1) poor generalization performance of modeling with a single downstream task, (2) uneven distribution of user behavior over time, and severe sparsity in many long-tail apps. In this paper, we propose a custom model (GPTCN) by which we overcome these challenges and optimize on top of it, achieving the goal of reducing manual effort and improving performance. Experimental results show that GPTCN outperforms the state-of-the-art general vector generation models in ACC, providing a more comprehensive, robust and general user representation model.