Research Article
A Multi-task Learning Framework with Features Based on Behavioral Pattern Conversation
@INPROCEEDINGS{10.1007/978-3-031-63992-0_17, author={Bo Tang and Nan Wang and Jinbao Li and Zhonghui Shen}, title={A Multi-task Learning Framework with Features Based on Behavioral Pattern Conversation}, proceedings={Mobile and Ubiquitous Systems: Computing, Networking and Services. 20th EAI International Conference, MobiQuitous 2023, Melbourne, VIC, Australia, November 14--17, 2023, Proceedings, Part II}, proceedings_a={MOBIQUITOUS PART 2}, year={2024}, month={7}, keywords={Multi-task Learning Behavioral Pattern Conversation Recommendation System}, doi={10.1007/978-3-031-63992-0_17} }- Bo Tang
Nan Wang
Jinbao Li
Zhonghui Shen
Year: 2024
A Multi-task Learning Framework with Features Based on Behavioral Pattern Conversation
MOBIQUITOUS PART 2
Springer
DOI: 10.1007/978-3-031-63992-0_17
Abstract
Deep neural network based multi-task learning has been widely successful in many real-world large scale applications, such as recommendation systems. A large number of commodity transaction data from e-commerce platforms show that users often go through a series of behavioral transitions such as impressed(\rightarrow )click(\rightarrow )add shopping cart before finally forming a purchase behavior, and only very few users click and then make a purchase directly. This phenomenon precisely follows the objective reality of power-law distribution. Based on this, this paper proposes a Multi-task learning framework with features based on Behavioral Pattern Conversation (BPCM). A feature tower model based on attribute information is constructed in the framework, which is able to control the fusion and screening process of features adaptively through the designed novel gate complementary mechanism. In addition, we designed several submodules with behavioral pattern conversation (BPC) applied to multi-task learning. The class of modules is not only able to adaptively model the sequentiality and dependencies between behavioral task transitions through information transfer, but also to effectively control the amount of information transferred between different tasks. Adequate experiments show that our BPCM obtains higher performance compared to more current advanced multi-task learning frameworks.
