DARec: Deep Domain Adaptation for Cross-Domain Recommendation via Transferring Rating Patterns

Feng Yuan, Lina Yao, Boualem Benatallah

Introduction

Recommender systems frequently encounter the "cold-start" problem and data sparsity, where insufficient interaction data hinders accurate predictions. Cross-Domain Recommendation (CDR) addresses this by leveraging knowledge from a related source domain to improve performance in a target domain. However, existing CDR methods often rely heavily on auxiliary side information (e.g., item descriptions or images) and utilize linear models like Matrix Factorization, which may fail to capture complex, non-linear user-item interaction patterns.

DARec (Deep Domain Adaptation for Cross-Domain Recommendation) is a novel framework designed to transfer "rating patterns" directly from sparse rating matrices without requiring auxiliary data. By employing adversarial domain adaptation, DARec bridges the gap between different product categories using deep learning.

The DARec Framework

DARec integrates AutoRec and Domain Adversarial Neural Networks (DANN) to extract and transfer domain-invariant features.

The architecture consists of four primary components:

  1. Embedding Generation: Utilizing AutoRec, the model compresses sparse rating matrices into dense, low-dimensional user or item embeddings.
  2. Rating Pattern Extractor: These embeddings are processed to identify underlying behavioral patterns.
  3. Adversarial Transfer: A Domain Classifier, coupled with a Gradient Reversal Layer (GRL), is used to ensure the extracted patterns are domain-invariant. The GRL "tricks" the extractor into learning features that make source and target domains indistinguishable to the classifier.
  4. Rating Predictor: Finally, these invariant features are used to predict numerical ratings in the target domain.

The authors propose two variants: U-DARec, which focuses on shared user patterns, and I-DARec, which focuses on distinctive item patterns.

Experimental Evaluation

The framework was evaluated using four cross-domain pairs from the Amazon public dataset:

  • Office Products $\rightarrow$ Movies and TV
  • Sports and Outdoors $\rightarrow$ CDs and Vinyl
  • Android Apps $\rightarrow$ Video Games
  • Toys and Games $\rightarrow$ Automotive

These datasets feature fully aligned users but distinct item sets, with extreme sparsity (>99.8% missing ratings).

Key Results and Analysis

DARec consistently outperformed state-of-the-art baselines, including traditional Matrix Factorization variants (CMF, CDTF) and deep learning models (CoNet).

As shown in the experimental results, both U-DARec and I-DARec variants generally achieve lower RMSE than their single-domain AutoRec counterparts across various embedding sizes. Notably, I-DARec often provides the most significant accuracy gains. For instance, in the "Toys & Automotive" pair, I-DARec achieved an RMSE of 2.0723, significantly outperforming CoNet (2.1356).

The results highlight several key insights:

  • Sufficiency of Ratings: Structural patterns within raw rating matrices contain sufficient signal for effective cross-domain transfer, negating the absolute necessity for side information.
  • Adversarial Efficacy: Adversarial learning successfully identifies a "common language" between disparate product categories.
  • Non-linear Advantage: The superiority over linear models suggests that user rating logic is inherently complex and better captured by deep neural architectures.

Conclusion and Future Work

DARec demonstrates that deep domain adaptation can effectively mitigate data sparsity in recommendation systems by transferring rating patterns. While the current model requires user overlap, it sets a strong foundation for future work, which may include integrating content-based features (text/images) to further enhance domain-invariant representation learning.

Summary updated

BibTeX
@inproceedings{yuan2019darec,
  author = {Feng Yuan and Lina Yao and Boualem Benatallah},
  title = {{DARec: Deep Domain Adaptation for Cross-Domain Recommendation via Transferring Rating Patterns}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/587},
}
Figure