PaperlayerView exact revision on arXiv

Source-linked AI summary

DDTCDR: Deep Dual Transfer Cross Domain Recommendation

Pan Li, Alexander Tuzhilin

arXiv:1910.05189v1cs.IRcs.LG

TL;DR

Cross-domain recommenders have not adequately modeled bidirectional latent user–item relations or user and item features. DDTCDR combines iterative dual transfer, latent orthogonal mappings, and autoencoder-based feature embeddings; the paper reports consistent gains over state-of-the-art and classical transfer-learning baselines across movie, book, and music domains.

  • Problem

    Existing cross-domain models mainly use ratings while omitting bidirectional latent user–item relations and explicit user and item features.

  • Method

    DDTCDR iteratively transfers latent preferences between domains using latent orthogonal mappings and autoencoder-derived user and item feature embeddings.

  • Results

    The model consistently and significantly outperforms state-of-the-art baselines and classical transfer-learning approaches across movie, book, and music domains.

  • Takeaways & Limitations

    Bidirectional latent transfer improves recommendation performance for both domains simultaneously while the model can extend to multiple-domain recommendation applications.

  • Takeaways & Limitations

    Improvements are relatively greater for the book and movie domains than for music, possibly because their datasets are significantly larger; this requires further study.

Abstract

from arXiv · show

Cross domain recommender systems have been increasingly valuable for helping consumers identify the most satisfying items from different categories. However, previously proposed cross-domain models did not take into account bidirectional latent relations between users and items. In addition, they do not explicitly model information of user and item features, while utilizing only user ratings information for recommendations. To address these concerns, in this paper we propose a novel approach to cross-domain recommendations based on the mechanism of dual learning that transfers information between two related domains in an iterative manner until the learning process stabilizes. We develop a novel latent orthogonal mapping to extract user preferences over multiple domains while preserving relations between users across different latent spaces. Combining with autoencoder approach to extract the latent essence of feature information, we propose Deep Dual Transfer Cross Domain Recommendation (DDTCDR) model to provide recommendations in respective domains. We test the proposed method on a large dataset containing three domains of movies, book and music items and demonstrate that it consistently and significantly outperforms several state-of-the-art baselines and also classical transfer learning approaches.

1 INTRODUCTION

The paper targets cross-domain recommendation limits involving sparse data, one-way transfer, latent user–item relations, and omitted feature information. DDTCDR combines bidirectional dual transfer, latent orthogonal mappings, and feature embeddings, with reported gains across domains and settings.

  • Motivation: Cross-domain recommendation addresses sparse transaction data by transferring user preferences between related domains.The paper gives movie-to-novel recommendation as an example of cross-domain transfer.
  • Research gap: Earlier models mainly transfer explicit information unidirectionally and omit latent user–item relations and user and item features.These limitations motivate combining dual transfer learning with latent embedding methods.
  • Approach: DDTCDR learns latent orthogonal mappings across domains while leveraging user preferences from all domains.The mapping is intended to preserve similarities between user preferences and support efficient inverse mapping.
  • Approach: The model transfers latent feature and preference representations rather than only explicit information, capturing latent interactions while modeling feature information.This design uses latent representations to address the limitations of explicit-information transfer.
  • Findings: Dual transfer enables bidirectional preference transfer that improves recommendation performance in both domains simultaneously over time.The paper reports empirical improvements across multiple domains and experimental settings.

2 RELATED WORK

Prior work uses cross-domain and transfer-learning methods to mitigate sparsity and cold start, but commonly improves only the target domain and omits explicit feature modeling. The paper positions dual transfer combined with autoencoding as a response to these gaps.

  • Cross-domain recommendation: Cross-domain recommendation methods use auxiliary-domain interaction information to address data sparsity.Examples include CMF, CDCF, CDFM, Canonical Correlation Analysis, and Dual Regularization.
  • Transfer learning: Transfer-learning methods learn preferences in a source domain and transfer them to a target domain, helping address cold start and improve recommendation performance.These methods assume shared knowledge that can be represented in a common latent feature space.
  • Research gap: Existing models do not improve recommendation performance in both domains simultaneously and do not explicitly model user and item features.The paper identifies these as limitations of prior cross-domain transfer approaches.
  • Proposed direction: The paper proposes combining a novel dual transfer mechanism with an autoencoder to overcome these issues and improve recommendation performance.The combination is presented as a response to the missing joint treatment of bidirectional transfer and feature information.
  • Dual transfer learning: Dual transfer learning simultaneously learns marginal and conditional distributions to exploit duality for mutual reinforcement.Earlier applications reported success in areas including machine translation.
  • Deep recommendation: Deep recommendation methods use embeddings and autoencoders to extract latent user–item interactions and better model user preferences.The related work connects these methods with the paper’s motivation for combining deep learning and dual transfer.

3 METHOD

DDTCDR combines autoencoded user and item feature embeddings with latent orthogonal mappings and iterative bidirectional transfer across domains. Its dual models estimate ratings from within-domain and cross-domain preferences, with convergence analyzed under stated assumptions and extension to multiple domains.

  • 3.1 Feature Embeddings: DDTCDR constructs feature embeddings with autoencoders before transferring them across domains through latent orthogonal mappings.The autoencoder transforms heterogeneous, discrete user and item features into continuous latent embeddings, while the mapping transfers preferences between domains.
  • 3.3 Deep Dual Transfer Learning: The model estimates ratings by combining within-domain and cross-domain user preferences, then jointly updates recommendation models and orthogonal mappings through backpropagation.The transfer rate α controls the relative contribution of the two preference components; the framework updates both domain models iteratively.
  • 3.2 Latent Orthogonal Mapping: Latent orthogonal mappings transfer user preferences while preserving similarities between users across domains and supporting efficient inverse mapping.The method assumes that users with similar preferences in one domain should have similar preferences in another.
  • 3.3 Deep Dual Transfer Learning: The dual transfer mechanism improves recommendations in both domains simultaneously by transferring user preferences bidirectionally during iterative learning.Within-domain preferences capture same-domain interactions, while cross-domain preferences use mapped information from the other domain.
  • 3.4 Convergence Analysis: The iterative optimization of dual matrix factorization for rating matrices is theoretically guaranteed to converge under the stated proposition, but convergence remains unclear for the neural matrix-factorization model.The paper tests the convergence hypothesis empirically because the theoretical guarantee is established for the dual matrix-factorization setting.
  • 3.5 Extension to Multiple Domains: DDTCDR extends beyond two domains by combining within-domain and cross-domain estimates through latent orthogonal transfer matrices for each domain pair.The proposed formulation is described as capable of providing recommendations for multiple-domain applications.

4 EXPERIMENT

The experiments evaluate DDTCDR on a large multi-domain recommendation dataset using cross-validation, standard recommendation metrics, and comparisons with several baseline models. They also examine user and item features and tune model hyperparameters.

  • Experimental design: Experiments compare cross-domain recommendation accuracy with state-of-the-art methods and study the sensitivity of DDTCDR's hyperparameters.The evaluation uses record-stratified 5-fold cross-validation and RMSE, MAE, Precision, and Recall.
  • Dataset: The dataset comes from a European online recommendation service where users rate and review items across independent domain sub-sites.The selected subset contains the three largest domains, although the supplied passage truncates before naming them.
  • User and item features: User features are derived from answers to eight popular questions covering preferences, recent events, lifestyles, and demographics.The platform also collects interaction records between users and items; the supplied passage does not fully describe the item-feature construction.
  • User and item features: Table 3 reports statistics of the feature sets used in the proposed model.The supplied caption identifies the table's subject but does not specify its row or column encoding.
  • Baseline models: Baselines include CCFNet, CDFM, CoNet, NCF, and CMF, representing content-boosted filtering, factorization machines, cross networks, neural collaborative filtering, and collective matrix factorization.NCF is trained separately for each domain, whereas CMF shares parameters among factors when users participate in multiple domains.
  • Implementation settings: The model uses Bayesian Optimization to set α=0.03, one-layer MLP encoders and decoders, and feature embeddings of size 8.Baseline methods use the same applicable hyperparameters as the proposed model.

5 RESULTS

DDTCDR is evaluated across three domain pairs and consistently outperforms baselines, while its gains vary by domain and its convergence is assessed empirically.

  • Cross-Domain Recommendation Performance: DDTCDR outperforms the second-best baselines across the Book-Movie, Book-Music, and Movie-Music domain pairs.The comparison uses RMSE, MAE, Precision, and Recall; Table 4 covers Book-Movie, Table 5 Book-Music, and Table 6 Movie-Music.
  • Cross-Domain Recommendation Performance: 0.2213 and 0.2213 RMSE measures for Book-Movie domains outperform the second-best baselines by 3.98% and 2.44%.The corresponding MAE measures are 0.1708 and 0.1704, with improvements of 9.54% and 9.80%.
  • Cross-Domain Recommendation Performance: 4.07%, 8.87%, 2.14%, and 4.74% are the reported improvements for Book-Music, versus 3.75%, 9.77%, 1.89%, and 4.24% for Movie-Music.These improvements are reported relative to the second-best baselines.
  • Cross-Domain Recommendation Performance: Improvements are relatively greater for the book and movie domains than for the music domain.The paper suggests the larger book and movie datasets as one possible reason and identifies further investigation as future work.
  • Convergence: DDTCDR stabilizes quickly and significantly outperforms NCF after 10 epochs.The model is trained iteratively for 100 epochs until the loss change is less than 1e-5, with training loss plotted for three domain pairs.
  • Hyperparameter Sensitivity: Cross-domain recommendation results remain consistently better than single-domain results for positive, small transfer-rate α values, despite performance fluctuation.The experiments vary hyperparameter settings to assess sensitivity to the transfer rate.

6 CONCLUSION

DDTCDR improves cross-domain recommendation by transferring latent information bidirectionally through embeddings and iterating until models stabilize. The approach uses latent orthogonal mappings to preserve user-preference similarity and is currently framed for domain pairs, with broader multi-domain transfer left for future work.

  • 6 CONCLUSION: DDTCDR transfers latent information between domains through embeddings and iterates the transfer loop until both models stabilize.The paper also states that convergence is guaranteed under certain conditions and empirically validated across experimental settings.
  • 6 CONCLUSION: The reported training-loss plots cover book, movie, and music domain pairs across Figures 2, 3, and 4.
  • 6 CONCLUSION: The model enables bidirectional training that improves performance measures for both domains simultaneously.
  • 6 CONCLUSION: Its latent orthogonal mapping preserves user-preference similarity and computes the inverse mapping efficiently.
  • 6 CONCLUSION: Extending the dual-learning mechanism from domain pairs to multiple domains remains future work.
Loading 1910.05189v1…