Hierarchical Projection Enhanced Multi-behavior Recommendation.
Chang Meng,Hengyu Zhang,Wei Guo,Huifeng Guo,Haotian Liu,Yingxue Zhang,Hongkun Zheng,Ruiming Tang,Xiu Li,Rui Zhang
DOI: https://doi.org/10.1145/3580305.3599838
2023-01-01
Abstract:Various types of user behaviors are recorded in most real-world recommendation scenarios. To fully utilize the multi-behavior information, the exploration of multiplex interaction among them is essential. Many multi-task learning based multi-behavior methods are proposed recently to use multiple types of supervision signals and perform information transfer among them. Despite the great successes, these methods fail to design prediction tasks comprehensively, leading to insufficient utilization of multi-behavior correlative information. Besides, these methods are either based on the weighting of expert information extracted from the coupled input or modeling of information transfer between multiple behavior levels through task-specific extractors, which are usually accompanied by negative transfer phenomenon 1 . To address the above problems, we propose a multi-behavior recommendation framework, called Hierarchical Projection Enhanced Multi-behavior Recommendation (HPMR). The key module, Projection-based Transfer Network (PTN), uses the projection mechanism to "explicitly" model the correlations of upstream and downstream behaviors, refines the upstream behavior representations, and fully uses the refined representations to enhance the learning of downstream tasks. Offline experiments on public and industrial datasets and online A/B test further verify the effectiveness of HPMR in modeling the associations from upstream to downstream and alleviating the negative transfer. The source code and datasets are available at https://github.com/MC-CV/HPMR.