Multi-domain Recommendation with Embedding Disentangling and Domain Alignment

By W. Ning et al

Published on Oct. 21, 2025

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

Table of Contents

1. Introduction
2. Problem Definition
3. ED Recommender
4. Intra-domain and Inter-domain Models

Summary

Multi-domain recommendation (MDR) aims to provide recommendations for different domains with overlapping users/items. Existing MDR models face challenges in disentangling knowledge across domains and transferring knowledge with small overlaps. The proposed EDDA method tackles these challenges with embedding disentangling and domain alignment. EDDA outperforms baselines on real datasets. The ED architecture disentangles inter-domain and intra-domain knowledge, while intra-domain and inter-domain models learn separately. Intra-domain model uses a graph-based network for each domain, while inter-domain model learns across all domains. Training involves Bayesian Personalized Ranking loss. Results show clear separation between intra-domain and inter-domain embeddings. EDDA introduces a domain alignment strategy based on random walks to enhance knowledge sharing across domains. The paper contributes by proposing an effective MDR method and conducting extensive experiments.