Source-linked AI summary
DARec: Deep Domain Adaptation for Cross-Domain Recommendation via Transferring Rating Patterns
Feng Yuan, Lina Yao, Boualem Benatallah
TL;DR
Cross-domain recommendation needs to transfer useful knowledge despite sparse ratings, while many deep methods rely on auxiliary content and rarely exploit rating-matrix patterns directly. DARec uses domain adaptation to extract and transfer shared patterns from rating matrices only, and the authors report superior performance to selected single-domain and cross-domain baselines on public datasets. The method is evaluated under an aligned-user setting, which bounds its stated scope.
Problem
Cross-domain recommendation methods often learn transferable features from auxiliary content, while direct transfer of patterns in sparse rating matrices remains underexplored.
Method
DARec uses deep neural networks with a rating pattern extractor, domain classifier, and rating predictor to transfer shared patterns between source and target rating matrices.
Results
DARec is reported as superior to selected state-of-the-art single-domain rating-prediction baselines, while I-DARec exceeds selected state-of-the-art deep cross-domain methods on public datasets.
Takeaways & Limitations
Rating matrices alone can support DARec’s extraction and transfer of shared rating patterns for cross-domain recommendation in the evaluated settings.
Takeaways & Limitations
The paper assumes explicit feedback and the same set of users in the source and target domains.
Abstract
from arXiv · showhide
Cross-domain recommendation has long been one of the major topics in recommender systems. Recently, various deep models have been proposed to transfer the learned knowledge across domains, but most of them focus on extracting abstract transferable features from auxilliary contents, e.g., images and review texts, and the patterns in the rating matrix itself is rarely touched. In this work, inspired by the concept of domain adaptation, we proposed a deep domain adaptation model (DARec) that is capable of extracting and transferring patterns from rating matrices {\em only} without relying on any auxillary information. We empirically demonstrate on public datasets that our method achieves the best performance among several state-of-the-art alternative cross-domain recommendation models.
1 Introduction
Cross-domain recommendation addresses sparse-feedback challenges by transferring knowledge across related domains, but prior deep approaches largely rely on auxiliary content rather than rating-matrix patterns. DARec applies domain adaptation to extract and transfer shared rating patterns from rating matrices only, with public-dataset results surpassing selected baselines and cross-domain methods.
- Motivation: Cross-domain recommendation leverages information from multiple domains to improve recommendations in a target domain despite limited feedback.Users and items can be grouped into domains in large real-world datasets, making CDR a practical approach.
- Research gap: Most prior deep cross-domain methods transfer knowledge from auxiliary contents, while direct extraction of patterns from sparse rating matrices remains rarely studied.Earlier clustering and matrix-factorization methods also struggle with nonlinear rating patterns and commonly require dense source-domain matrices.
- Approach: DARec uses domain adaptation to learn shared rating patterns across source and target domains with a rating pattern extractor, domain classifier, and rating predictor.The model is inspired by domain-adversarial neural networks and addresses distribution differences between domains.
- Contributions: DARec extracts and transfers abstract rating patterns using only rating matrices, with U-DARec targeting shared user patterns and I-DARec separating distinctive item patterns.Adversarial training transfers shared patterns for aligned users, while item-based modeling separates patterns associated with unrelated items.
- Empirical findings: On public datasets, DARec outperforms selected state-of-the-art single-domain rating-prediction baselines and I-DARec exceeds selected state-of-the-art deep cross-domain methods.The reported comparisons attribute the gains to transferred source-domain information and evaluate both single-domain and cross-domain alternatives.
2 Related Work
Prior cross-domain recommendation methods use clustering, matrix factorization, or deep learning, often relying on auxiliary content or dense data. DARec instead uses deep neural networks and domain adaptation to transfer rating-matrix patterns without auxiliary features or dense source data.
- Existing approaches: Collaborative-filtering CDR methods evolved from neighborhood approaches toward matrix factorization, while content-based methods establish links through auxiliary user or item information.Content-based approaches can avoid user or item overlap and relatively dense rating matrices by using auxiliary features.
- Existing approaches: Auxiliary-content methods learn shared domain patterns from features such as users, news, apps, or items, emphasizing rating patterns less directly.These methods include multi-view and domain-adaptation frameworks that combine or extract features from auxiliary data.
- DARec distinction: DARec uses DNNs to learn linear and nonlinear shared rating patterns directly from raw sparse rating matrices without auxiliary user or item features.The approach also does not require dense source-domain data for extracting transferable knowledge.
- DARec distinction: DARec combines domain adaptation with deep rating-pattern learning to alleviate distribution mismatch between domains rather than modeling domain differences only through fully connected networks.This design distinguishes it from CoNet in the authors’ comparison of related methods.
3 Proposed Method
DARec formulates shared-user cross-domain recommendation by reconstructing rating vectors, extracting latent rating patterns, and adapting them across source and target domains. It uses adversarial domain classification for shared user patterns and a separate item-based objective for domain-exclusive patterns.
- Problem Definition: The problem assumes explicit feedback, shared users across source and target domains, and different item sets, with recommendation based on predicting unknown target-domain ratings.The task leverages the source rating matrix to predict target-domain ratings for users.
- AutoRec Representations: AutoRec maps each partially observed user or item rating vector into a low-dimensional latent space and reconstructs the vector to generate missing ratings.U-AutoRec represents users, while I-AutoRec represents items.
- User-Based DARec: U-DARec interleaves source- and target-domain user embeddings before applying the modified DANN to extract shared rating patterns.The latent factors are computed from trained AutoRec parameters and then fed into the domain-adaptation network.
- User-Based DARec: The rating pattern extractor, predictor, and domain classifier are trained jointly so predictors reconstruct ratings while extracted features become less domain-discriminative.The extractor maximizes domain-classifier loss, whereas the classifier and predictor minimize their respective losses.
- Item-Based DARec: I-DARec extracts domain-exclusive item patterns because the source and target item sets do not overlap, then directly optimizes the resulting objective without a GRL layer.The item patterns are separated using the rating predictor, and SGD-like algorithms optimize Equation 7 directly.
4 Experiments
DARec is evaluated on extremely sparse Amazon cross-domain datasets using rating matrices and compared with single-domain and cross-domain baselines. The experiments show gains from domain adaptation, with I-DARec achieving the strongest baseline performance.
- Datasets: The evaluation uses Amazon item categories as domains, selecting users with at least five ratings; both domains have at least 99.8% unobserved ratings.
- Single-domain baselines: AutoRec, PMF, RBM, and CF-NADE are used as single-domain rating-prediction comparators, with hyperparameters tuned for best performance.
- Transfer effectiveness: U-DARec and I-DARec both improve over their corresponding AutoRec variants by transferring shared or distinct rating patterns.
- Transfer effectiveness: In the Office Products–Movies & TV dataset, U-DARec improves RMSE by 3.66% over U-AutoRec, while the Sports & Outdoors–CDs & Vinyl pair improves by 4.30%.
- Embedding size: Embedding size has an optimum: U-DARec reaches minimum RMSE near 800 dimensions and I-DARec near 400 on the Office Products–Movies & TV pair.Too-small embeddings hinder pattern extraction, whereas oversized embeddings introduce noisy latent details.
- Cross-domain baselines: Across cross-domain baselines, deep methods outperform non-deep methods on datasets with over 99.8% sparsity, and I-DARec achieves the best performance among the baselines.For Office Products–Movies & TV with 90% training data, I-DARec improves RMSE by 3.45% over FM-CDCF.
5 Conclusions
The paper presents DARec for aligned-user cross-domain recommendation using only rating matrices. It extracts shared user patterns and distinct item patterns through domain adaptation, and reports strong knowledge transfer and recommendation performance.
- DARec addresses aligned-user cross-domain recommendation without using contextual or auxiliary information, relying only on rating matrices.
- AutoRec produces user embeddings, which are interleaved across source and target domains and processed by a domain-adaptation network.
- A domain classifier extracts shared user rating patterns, while a rating predictor separates estimated ratings for each domain.
- The extended model uses item ratings as inputs to extract distinct features for items in the two domains.
- Comparisons with single-domain recommendation methods demonstrate knowledge transfer, and the proposed model achieves the best reported performance.