Cross-Domain Recommendation Task

A Cross-Domain Recommendation Task is a Machine Learning Task that involves leveraging user and item data from one or more source domains to improve recommendation performance in a target domain.

• AKA: Cross-Domain Recommender Task, Multi-Domain Recommendation Task.
• Context:
  • Task Input: User-item interaction data from both source and target domains.
    • Optional Input: User profiles, item metadata, contextual information (e.g., timestamps, location).
  • Task Output: Personalized item recommendations in the target domain.
  • Task Performance Measures: Metrics such as Precision, Recall, F1-score, NDCG, and RMSE.
  • Task Objective: Enhance recommendation accuracy in the target domain by utilizing knowledge from source domains.
  • It can be systematically solved and automated by a Cross-Domain Recommendation System.
  • It can address challenges like data sparsity and cold-start problems by transferring knowledge from richer domains to sparser ones.
  • It can involve various scenarios, including single-target (improving recommendations in one domain) and dual-target (simultaneously enhancing recommendations in multiple domains) settings.
  • It can utilize techniques such as collaborative filtering, matrix factorization, transfer learning, and graph neural networks to model cross-domain relationships.
  • It can be applied in diverse fields like e-commerce, where user behavior in one category (e.g., electronics) informs recommendations in another (e.g., books).
  • It can consider different levels of domain relationships, such as content-level (shared attributes), user-level (common users), and item-level (shared items).
  • ...
• Example(s):
  • Recommending books to users based on their movie viewing history.
  • Suggesting new music tracks to users by analyzing their podcast listening patterns.
  • Providing travel destination recommendations by leveraging users' restaurant review data.
  • ...
• Counter-Example(s):
  • Single-Domain Recommendation Task, which focuses solely on recommendations within one domain without leveraging external domain data.
  • Cold-Start Recommendation Task, which specifically addresses recommendations for new users or items without prior interactions, not necessarily involving multiple domains.
  • Cross-Domain Recommendation Benchmarking Task, which is designed to evaluate model performance rather than improve cross-domain recommendations.
  • ...
• See: Domain-Specific Text Understanding Task, Cross-Domain Transfer Learning Task, Automated Domain-Specific Writing Task, Targeted Concept Simplification Task, Transfer Learning, Collaborative Filtering, Matrix Factorization, Graph Neural Networks.

---

---

References

2023

• (Lin et al., 2023) ⇒ Xuan Lin, Jianliang Gao, Zhenyu He, Yanyan Liu, & Binghong Chen. (2023). "Dual-Target Cross-Domain Recommendation with Contrastive Learning". arXiv Preprint.
  • QUOTE: We propose DCCL, a contrastive learning framework for dual-target cross-domain recommendation that simultaneously improves performance in both source and target domains. DCCL achieves 12.7% and 9.3% MAE reduction on Amazon and MovieLens datasets compared to single-domain baselines by modeling domain-shared user preferences through contrastive alignment.

2022

• (Dacrema et al., 2022) ⇒ Maurizio Ferrari Dacrema, Iván Cantador, Ignacio Fernández-Tobías, Shlomo Berkovsky, & Paolo Cremonesi. (2022). "Design and Evaluation of Cross-Domain Recommender Systems". In: Recommender Systems Handbook.
  • QUOTE: Cross-domain recommender systems enhance recommendations by exploiting knowledge transfer between source and target domains through three principal technique categories: aggregating user preferences, linking domain features, and transferring latent patterns. Experimental results show matrix co-factorization methods improve RMSE by 18.4% in cold-start scenarios compared to single-domain approaches.

2021

• (Zhu et al., 2021) ⇒ Feng Zhu, Yan Wang, Chaochao Chen, Jun Zhou, Longfei Li, & Guanfeng Liu. (2021). "Cross-Domain Recommendation: Challenges, Progress, and Prospects". In: Proceedings of IJCAI 2021.
  • QUOTE: This survey categorizes cross-domain recommendation approaches into four taxonomies: single-target, single-target multi-domain, dual-target, and multi-target CDR. Analysis of 127 papers reveals deep transfer learning methods achieve 23.7% higher NDCG@10 than traditional matrix factorization in partial user overlap scenarios.

2014

• (Cantador & Cremonesi, 2014) ⇒ Iván Cantador & Paolo Cremonesi. (2014). "Cross-domain Recommender Systems". In: Proceedings of ACM RecSys 2014.
  • QUOTE: Cross-domain recommendation addresses data sparsity and cold-start problems by leveraging user preference overlap across domains. The tutorial identifies three domain linkage scenarios: attribute-level (e.g., movie genres), type-level (e.g., books vs. movies), and system-level (e.g., Netflix vs. MovieLens), with system-level sharing showing 15% precision improvement in empirical studies.

---