Source-linked AI summary
Cross-Domain Recommendation: Challenges, Progress, and Prospects
Feng Zhu, Yan Wang, Chaochao Chen, Jun Zhou, Longfei Li, Guanfeng Liu
TL;DR
Recommender systems face persistent data sparsity, while existing CDR research lacks a systematic review. This paper surveys four CDR scopes, their challenges and taxonomies, research progress, and future directions, concluding that multi-target CDR remains challenging because existing approaches do not address negative transfer.
Problem
Existing collaborative-filtering recommender systems suffer from data sparsity, and CDR lacks a systematic review covering its approaches, challenges, and research progress.
Method
The paper surveys single-target CDR, multi-domain recommendation, dual-target CDR, and multi-target CDR, presenting their definitions, challenges, taxonomies, research progress, and prospects.
Results
The survey provides a full-view categorization of CDR approaches and summarizes representative research efforts and trends across four recommendation scopes.
Takeaways & Limitations
The survey identifies promising directions for CDR research and is intended to facilitate future research in the community.
Takeaways & Limitations
Multi-target CDR remains challenging because existing approaches do not consider the negative transfer problem.
Abstract
from arXiv · showhide
To address the long-standing data sparsity problem in recommender systems (RSs), cross-domain recommendation (CDR) has been proposed to leverage the relatively richer information from a richer domain to improve the recommendation performance in a sparser domain. Although CDR has been extensively studied in recent years, there is a lack of a systematic review of the existing CDR approaches. To fill this gap, in this paper, we provide a comprehensive review of existing CDR approaches, including challenges, research progress, and future directions. Specifically, we first summarize existing CDR approaches into four types, including single-target CDR, multi-domain recommendation, dual-target CDR, and multi-target CDR. We then present the definitions and challenges of these CDR approaches. Next, we propose a full-view categorization and new taxonomies on these approaches and report their research progress in detail. In the end, we share several promising research directions in CDR.
1 Introduction
Cross-domain recommendation addresses recommender-system data sparsity by transferring richer-domain information to improve recommendations in a sparser target domain. This survey reviews CDR challenges, approaches, research progress, and future directions.
- Motivation: Data sparsity, especially for new users or items, can cause over-fitting and reduce recommendation accuracy in collaborative-filtering systems.CDR addresses this problem by leveraging information from a richer source domain for a sparser target domain.
- Motivation: CDR transfers user, item, preference, or interaction information from a richer source domain to improve recommendation accuracy in a sparser target domain.Examples of transferable information include user or item features, thumbs-up signals, tags, reviews, and ratings.
- CDR Scope: CDR domains may be linked through content-level relevance, user-level relevance, or item-level relevance, including settings without common users or items.The survey distinguishes shared content, common users with different item levels, and common items with different users.
- Approach Taxonomy: Existing CDR approaches are classified as content-based transfer, embedding-based transfer, and rating pattern-based transfer.These categories organize conventional approaches according to how information is transferred across domains.
- CDR Scope: Conventional single-target CDR uses auxiliary information from a richer domain to help a sparser domain, while newer dual-target and multi-target settings improve multiple domains.The survey identifies dual-target and multi-target CDR as newer research directions motivated by domains having different types of useful information.
- Survey Contributions: The survey analyzes CDR challenges, classifies approaches, compares their contributions and differences, reviews research progress, and outlines promising future directions.Its contributions include a developing-perspective overview and a comprehensive review of current CDR research.
2 Problems and Challenges
CDR is organized into four recommendation scenarios that differ in domain scale, overlap, and improvement targets. These scenarios face challenges involving cross-domain relations, representation quality, simultaneous improvement, and negative transfer.
- CDR scenarios include single-target CDR, multi-domain recommendation, dual-target CDR, and multi-target CDR.Their main differences concern domain scales, overlaps, and improvement targets.
- Single-Target CDR: Single-target CDR improves a sparser target domain by leveraging richer explicit, implicit, and side information from a source domain.The source and target domains may contain different user and item sets.
- Single-Target CDR: Single-target CDR requires suitable content-based relations and accurate user/item embeddings or rating patterns for cross-domain transfer.These representations support selecting similar entities and sharing information across domains.
- Multi-Domain Recommendation: Multi-domain recommendation uses multiple domains to recommend items from those domains to a selected user set while facing the same challenges as single-target CDR.The formulation targets corresponding recommendation accuracy across the multiple domains.
- Dual-Target CDR: Dual-target CDR improves recommendation accuracy in two domains simultaneously by leveraging their observed information and shared cross-domain entities.Relations may rely on common features, users, or items, while embedding optimization is particularly important.
- Multi-Target CDR: Multi-target CDR seeks simultaneous improvement across multiple domains but must avoid increasingly serious negative transfer as additional, especially sparser, domains participate.Transferred information may reduce recommendation performance in some domains, making negative-transfer avoidance a prerequisite.
3 Research Progress
The survey organizes CDR research by target scenarios, transfer strategies, and datasets, covering single-target, multi-domain, dual-target, and multi-target settings. It reviews how existing approaches transfer content, embeddings, rating patterns, or shared information, while highlighting unresolved challenges in multi-target CDR.
- Existing CDR research is categorized by target scenarios, target challenges, data categories, and technical perspectives, alongside summaries of popular datasets.
- Single-Target CDR: Single-target CDR approaches are grouped into content-based, embedding-based, and rating pattern-based transfer according to their transfer strategies.
- Single-Target CDR: Content-based transfer links domains through shared attributes, tags, semantic properties, text, metadata, or histories before transferring user or item information.
- Single-Target CDR: Embedding-based transfer maps or shares latent user/item factors using methods including multi-task learning, transfer learning, deep neural networks, and relational learning.
- Single-Target CDR: Rating pattern-based transfer learns users’ independent rating patterns in a source domain and transfers them to improve recommendation accuracy in the target domain.
- Dual-Target CDR: Dual-target CDR seeks simultaneous improvement in both domains through embedding combination, sharing, bidirectional transfer, and attention mechanisms.DTCDR combines multi-source information and common-user embeddings, while DDTCDR uses bidirectional latent orthogonal mapping.
- Multi-Target CDR: Multi-target CDR aims to address data sparsity across multiple domains, but few existing solutions consider the inevitable negative-transfer challenge.Shared heterogeneous graphs and source-domain auxiliary information have been used, yet the reviewed approaches do not address negative transfer.
- Datasets: The survey also summarizes commonly used CDR datasets and directs researchers to their corresponding citations and websites.
4 Research Prospects
The survey identifies heterogeneous and sequential CDR as promising directions. These directions reflect the mismatch between information types across domains and the challenge of modeling sequential user-item interactions across domains.
- Heterogeneous CDR: Heterogeneous CDR addresses the reality that different domains are rich in different information types, requiring techniques that identify bridges for transferring information.E-commerce may provide abundant user-item interactions, whereas social domains may provide plentiful user-user social data.
- Sequential CDR: Sequential CDR extends cross-domain recommendation toward modeling sequential dependencies in user-item interactions, including high-order, long-term, and noisy behavior.The survey describes sequential modeling as a challenge shared by CDR and conventional recommender systems.
5 Conclusion
The survey reviews four CDR scopes, defines their challenges, proposes comprehensive taxonomies, summarizes research trends, and identifies promising future prospects.
- The survey covers single-target CDR, multi-domain recommendation, dual-target CDR, and multi-target CDR.
- It presents definitions and challenges, develops full-view categorizations and new taxonomies, and details existing research progress.
- The survey lists promising CDR prospects and aims to facilitate future research in the community.