CATN: Cross-Domain Recommendation for Cold-Start Users via Aspect Transfer Network

Cheng Zhao; Chenliang Li; Rong Xiao; Hongbo Deng; Aixin Sun

Source-linked AI summary

CATN: Cross-Domain Recommendation for Cold-Start Users via Aspect Transfer Network

Cheng Zhao, Chenliang Li, Rong Xiao, Hongbo Deng, Aixin Sun

arXiv:2005.10549v2May 21, 2020cs.IRcs.LG

TL;DR

Cross-domain recommendation must transfer preferences for users lacking target-domain interactions, but existing pipelines do not explicitly model fine-grained semantic matching. CATN extracts review-derived aspects, learns their cross-domain correlations with attention, and uses auxiliary reviews in end-to-end training; it significantly outperforms existing models and provides fine-grained aspect connections.

Problem
Cross-domain recommendation seeks to transfer preferences from a source domain to a relevant target domain for cold-start users, while existing representation mapping cannot explicitly distinguish fine-grained semantic characteristics.
Method
CATN extracts aspects from user and item reviews, learns global cross-domain aspect correlations with attention, enhances representations with auxiliary reviews, and optimizes the model end to end.
Results
CATN significantly outperforms all baselines on cross-domain recommendations across different overlapping-user ratios, with more robust performance when overlap is scarce.
Takeaways & Limitations
CATN reveals user aspect connections across domains at fine granularity, supporting explainable recommendation for cold-start users.
Takeaways & Limitations
Existing comparative methods may require 120-core preprocessing, which does not fit real-world sparse recommender scenarios.

Abstract

from arXiv · show

In a large recommender system, the products (or items) could be in many different categories or domains. Given two relevant domains (e.g., Book and Movie), users may have interactions with items in one domain but not in the other domain. To the latter, these users are considered as cold-start users. How to effectively transfer users' preferences based on their interactions from one domain to the other relevant domain, is the key issue in cross-domain recommendation. Inspired by the advances made in review-based recommendation, we propose to model user preference transfer at aspect-level derived from reviews. To this end, we propose a cross-domain recommendation framework via aspect transfer network for cold-start users (named CATN). CATN is devised to extract multiple aspects for each user and each item from their review documents, and learn aspect correlations across domains with an attention mechanism. In addition, we further exploit auxiliary reviews from like-minded users to enhance a user's aspect representations. Then, an end-to-end optimization framework is utilized to strengthen the robustness of our model. On real-world datasets, the proposed CATN outperforms SOTA models significantly in terms of rating prediction accuracy. Further analysis shows that our model is able to reveal user aspect connections across domains at a fine level of granularity, making the recommendation explainable.

1 INTRODUCTION

Cross-domain recommendation addresses cold-start users by transferring preferences from a relevant source domain to a target domain. CATN performs this transfer at the aspect level using reviews and attention-based cross-domain correlations.

Cold-start users lack historical feedback in the target domain, although interactions in a relevant source domain can support recommendations there.
Existing methods map whole user-preference vectors between domains, which cannot explicitly distinguish fine-grained semantic characteristics.
CATN extracts multiple aspects from user and item review documents and learns their cross-domain correlations with attention.
The framework uses user and item documents formed from reviews, applying aspect-specific gates over convolutional layers for aspect extraction.
Across three pairs of real-world datasets, CATN significantly outperforms state-of-the-art alternatives and supports fine-grained semantic analysis.

2 RELATED WORK

Prior cross-domain methods transfer representations across domains, while aspect-based recommendation methods generally remain single-domain. CATN combines cross-domain transfer with aspect-level modeling to address cold-start recommendation through reviews.

Cross-Domain Recommendation: Cross-domain recommendation methods transfer knowledge through shared ratings, user factors, neural mappings, cross-stitch networks, item embeddings, or user neighborhoods.
Cross-Domain Recommendation: Many existing cross-domain solutions use rating records while overlooking reviews and other complementary information.
Cross-Domain Recommendation: Existing review-based transfer methods learn source and target representations before mapping overlapping users, limiting fine-grained semantic distinction and accumulating intermediate-step noise.
Aspect-Based Recommendation: Review-based recommender systems use reviews to address sparsity, while aspect-based systems model fine-grained relations between user preferences and item characteristics.
Aspect-Based Recommendation: External-toolkit approaches depend on extracted aspect and sentiment quality, whereas internal models automatically learn aspects but may use static representations.
Aspect-Based Recommendation: Single-domain aspect-based methods cannot handle target-domain cold-start users, motivating CATN’s cross-domain aspect transfer network.

3 THE CATN FRAMEWORK

CATN formulates cross-domain recommendation for cold-start users as aspect-level preference matching between source and target domains. It extracts aspects from review documents, enhances user representations with auxiliary reviews, learns cross-domain aspect correlations, and optimizes two reciprocal learning flows.

3.1 Problem Formulation: For a cold-start user with source-domain interactions but none in the target domain, CATN estimates ratings for target-domain items.
3.2 Overview of CATN: CATN uses overlapping users’ reviews and ratings from both domains to train two reciprocal prediction flows across source and target domains.One flow predicts target ratings from source-user and target-item documents; the other predicts source ratings from target-user and source-item documents.
3.3 Aspect Extraction: Aspect extraction applies convolution, aspect-specific gates, and attention guided by global domain-specific aspect representations to user and item documents.The gate selects latent features relevant to each aspect, while attention weights words for each aspect and produces aspect matrices for users and items.
3.4 Auxiliary Reviews Enhancement: Auxiliary documents merge reviews from non-overlapping like-minded users with matching rating scores, expanding training information when overlapping users are scarce.A second convolutional layer processes these potentially incoherent reviews before shared aspect extraction, and a gate updates the user’s aspect representation.
3.5 Cross-Domain Aspect Correlation Learning: The cross-domain correlation matrix weights preference transfer from source aspect p to target aspect q and uses LeakyReLU to support sparse correlations.

4 EXPERIMENTS

Experiments evaluate CATN on Amazon Book, Movie, and Music domains under simulated cold-start settings with varying overlap ratios. CATN significantly outperforms the baselines across scenarios and overlap settings, while showing robust performance when training overlap is scarce.

Datasets and protocol: Experiments use three Amazon domains—Book, Movie, and Music—and construct three paired cross-domain recommendation scenarios.Users with missing target-domain interactions are treated as cold-start users; 50% of overlapping users are reserved for validation and testing, with training overlap varied through η.
Experimental setup: MSE is the evaluation metric, computed on cold-start validation or test sets, with all methods tuned by grid search and evaluated over five runs.CATN uses n = 50 convolution filters, window size s = 3, dropout keep probability 0.8, learning rate 0.001, and searches latent dimension k and aspect number M.
Overall results: CATN significantly outperforms every baseline across all cross-domain scenarios and all tested ratios of overlapping users.Table 3 reports MSE comparisons and relative improvements over the best SOTA algorithm; reported improvements are statistically significant at the 0.05 level.
Baseline comparison: CMF consistently performs worst, while CDLFM improves over EMCDR but no compared method achieves the best result across evaluations.The authors attribute CMF’s weakness to joint-matrix factorization and describe CDLFM’s mapping improvements over EMCDR.
Robustness analysis: CATN is more robust than three-step optimization methods when the overlapping-user training ratio η is low, especially at 10% or 5%.The paper links this robustness to transferring cross-domain aspects rather than directly mapping user representations, reducing the impact of scarce overlap.

5 MODEL ANALYSIS

CATN analysis shows robustness to aspect-number choices, gains from global and separate aspect extraction plus auxiliary reviews, and competitive optimization efficiency. Case studies further show sparse, semantically coherent aspect transfers that support explainable recommendations.

5.1 Aspect Number Sensitivity: CATN remains robust across aspect-number settings because performance fluctuations are very small for fixed η values.Small M produces coarse aspects, whereas large M produces fine-grained aspects; attention weights matching aspects across domains.
5.1 Aspect Number Sensitivity: More overlapping users improve preference matching across domains and recommendation accuracy.η directly controls the number of overlapping users used to learn cross-domain preference matching.
5.2 Ablation Study: CATN-basic outperforms most baselines, while global-sharing representations, separate extraction parameters, and auxiliary reviews progressively improve performance.The integrated CATN variant performs best among the ablation variants, and auxiliary reviews help alleviate data sparsity.
5.3 Optimization Efficiency: CATN uses end-to-end optimization over overlapping-user ratings rather than the staged transfer process used by existing approaches.The paper motivates this design by noting that intermediate sub-optimal learning can accumulate and magnify noisy information, while prior staged training is time-consuming.
5.3 Optimization Efficiency: CATN reaches best validation performance in about 600s for Book →Movie at η = 50%, faster than EMCDR, CDLFM, DFM, and R-DFM.The comparison uses one Nvidia 1080 GPU; CMF and ANR are faster, at 300s and 400s respectively.
5.4 Explainability Analysis: The example analyses conclude that CATN provides reasonable aspect transfer supporting semantic explanation in cold-start recommendation.Aspect words and highlighted review sentences are used to inspect the most correlated source-target aspect pairs.
5.4 Explainability Analysis: Global aspect correlations are usually sparse, and examples show plausible transfers such as Book plot and scene to Movie content.The examples also connect movie plot preferences to soothing lute music and book story preferences to music praised for melody but criticized for lyrics.

6 CONCLUSION

The paper addresses review-based cross-domain recommendation for cold-start users by transferring multi-faceted preferences between domains. CATN uses end-to-end aspect-level correlation learning and outperforms existing models while supporting fine-grained explanations.

6 CONCLUSION: CATN targets effective and explainable transfer of review-derived user preferences from a source domain to a target domain for cold-start users.The framework focuses on multi-faceted preferences and matching only the aspects shared across domains.
6 CONCLUSION: CATN derives aspects from review documents and learns their cross-domain correlations through global aspect representations with attention.This replaces the existing sequence of separately learning domain representations and then learning a mapping.
6 CONCLUSION: CATN outperforms all existing models on cross-domain recommendation tasks.The conclusion reports this as the principal outcome of the proposed framework.
6 CONCLUSION: The model offers fine-grained user-preference connections across domains as an alternative basis for explainable recommendation.The paper identifies more effective modeling of cross-domain preference transfer as a direction for future study.