[Paper Review] Domain-to-Domain Translation Model for Recommender System

Linh Thuy Thi Nguyen, Tsukasa Ishigaki|arXiv (Cornell University)|Dec 15, 2018

Recommender Systems and Techniques19 references4 citations

TL;DR

This paper proposes a Domain-to-Domain Translation Model (D2D-TM) for multi-domain recommender systems that leverages Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Cycle-Consistency (CC) to jointly learn both shared (homogeneous) and domain-specific (divergent) features from user interaction histories. The model achieves state-of-the-art performance, outperforming existing methods by up to 129% in Recall@50 on real-world datasets.

ABSTRACT

Recently multi-domain recommender systems have received much attention from researchers because they can solve cold-start problem as well as support for cross-selling. However, when applying into multi-domain items, although algorithms specifically addressing a single domain have many difficulties in capturing the specific characteristics of each domain, multi-domain algorithms have less opportunity to obtain similar features among domains. Because both similarities and differences exist among domains, multi-domain models must capture both to achieve good performance. Other studies of multi-domain systems merely transfer knowledge from the source domain to the target domain, so the source domain usually comes from external factors such as the search query or social network, which is sometimes impossible to obtain. To handle the two problems, we propose a model that can extract both homogeneous and divergent features among domains and extract data in a domain can support for other domain equally: a so-called Domain-to-Domain Translation Model (D2D-TM). It is based on generative adversarial networks (GANs), Variational Autoencoders (VAEs), and Cycle-Consistency (CC) for weight-sharing. We use the user interaction history of each domain as input and extract latent features through a VAE-GAN-CC network. Experiments underscore the effectiveness of the proposed system over state-of-the-art methods by a large margin.

Motivation & Objective

• Address the challenge of capturing both shared and domain-specific features in multi-domain recommender systems.
• Overcome limitations of single-domain models that fail on cold-start and cross-selling scenarios.
• Eliminate reliance on external data sources like social networks or search queries for knowledge transfer.
• Enable bidirectional domain translation and end-to-end learning using only user interaction history.
• Develop a robust, scalable framework suitable for real-world deployment with minimal retraining.

Proposed method

• Integrate VAEs to learn latent representations of user interaction vectors in each domain.
• Employ GANs to enforce realism in generated click vectors and reduce VAE overfitting.
• Apply Cycle-Consistency (CC) with weight-sharing in high-level encoder and generator layers to enforce bidirectional translation between domains.
• Use shared latent space with tied weights in last encoder and first generator layers to model domain similarities.
• Train the model end-to-end using multinomial log-likelihood loss for better alignment with binary click data.
• Implement adversarial discriminators per domain to validate the realism of translated vectors.

Experimental results

Research questions

• RQ1 Can a unified model effectively learn both shared and divergent features across multiple domains in a recommender system?
• RQ2 How does joint learning via VAE-GAN-CC improve performance compared to isolated models or standard VAEs?
• RQ3 To what extent can the model generalize across domains without relying on external signals like social media or search queries?
• RQ4 How does the choice of reconstruction loss (e.g., multinomial log loss vs. L1/square loss) affect performance on click data?
• RQ5 Does bidirectional domain translation enhance recommendation accuracy compared to unidirectional transfer?

Key findings

• The D2D-TM model outperforms Multi-VAE by 44.8% in Recall@50 on the Health_Clothing dataset.
• It surpasses CDL by 17.9% to 129% in Recall@50 across different domain pairs, despite CDL’s hybrid design.
• The model improves by 8.1% over a VAE-only baseline in Health and Personal Care due to CC and GAN integration.
• Multinomial log-likelihood loss yields the best performance, outperforming L1, square, and standard log-likelihood losses on click data.
• The D2D-TM VAE_GAN variant achieves slightly better results than D2D-TM VAE_CC, indicating GAN’s strong regularization effect.
• The model maintains robust performance even with sparse user interactions, demonstrating scalability and low retraining needs.