Hamur: Hyper Adapter for Multi-Domain Recommendation

By Xiaopeng Li et al

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

1. Abstract
2. Introduction
3. Key Concepts
4. Methodology
5. Domain-Specific Adapter Cell
6. Domain Shared Hyper-Network
7. Integration with Backbone Model
8. Algorithm

Summary

The paper introduces Hamur, a novel model for Multi-Domain Recommendation (MDR), addressing the challenges of CTR prediction in diverse domains. It proposes a domain-specific adapter cell and a domain-shared hyper-network to dynamically generate adapter parameters. The adapter is integrated into existing MDR models for end-to-end training. The low-rank matrix decomposition method is utilized for computational efficiency. Experimental results demonstrate the effectiveness and scalability of Hamur in comparison to state-of-the-art methods.