PaperlayerView exact revision on arXiv

Source-linked AI summary

A Deep Framework for Cross-Domain and Cross-System Recommendations

Feng Zhu, Yan Wang, Chaochao Chen, Guanfeng Liu, Mehmet Orgun, Jia Wu

arXiv:2009.06215v1cs.LGcs.AI

TL;DR

CDR and CSR seek to alleviate recommender-system data sparsity by mapping information from richer source domains or systems to target settings, but complex latent-factor relationships make accurate mapping difficult. DCDCSR combines matrix factorization with a fully connected deep neural network and sparsity-guided training, and experiments on three real-world datasets show improved accuracy over state-of-the-art CDR and CSR approaches.

  • Problem

    CDR and CSR require accurate mapping of latent factors between source and target domains or systems, despite complex relationships and differing rating sparsity.

  • Method

    DCDCSR uses matrix-factorization models to generate latent factors, sparsity degrees to construct benchmark factors, and a fully connected deep neural network to map target factors.

  • Results

    DCDCSR outperforms state-of-the-art CDR and CSR approaches in recommendation accuracy across experiments on three real-world datasets.

  • Takeaways & Limitations

    Sparsity-guided deep mapping provides a framework for using source and target rating information in both cross-domain and cross-system recommendation.

Abstract

from arXiv · show

Cross-Domain Recommendation (CDR) and Cross-System Recommendations (CSR) are two of the promising solutions to address the long-standing data sparsity problem in recommender systems. They leverage the relatively richer information, e.g., ratings, from the source domain or system to improve the recommendation accuracy in the target domain or system. Therefore, finding an accurate mapping of the latent factors across domains or systems is crucial to enhancing recommendation accuracy. However, this is a very challenging task because of the complex relationships between the latent factors of the source and target domains or systems. To this end, in this paper, we propose a Deep framework for both Cross-Domain and Cross-System Recommendations, called DCDCSR, based on Matrix Factorization (MF) models and a fully connected Deep Neural Network (DNN). Specifically, DCDCSR first employs the MF models to generate user and item latent factors and then employs the DNN to map the latent factors across domains or systems. More importantly, we take into account the rating sparsity degrees of individual users and items in different domains or systems and use them to guide the DNN training process for utilizing the rating data more effectively. Extensive experiments conducted on three real-world datasets demonstrate that DCDCSR framework outperforms the state-of-the-art CDR and CSR approaches in terms of recommendation accuracy.

1 Introduction

CDR and CSR address recommender-system data sparsity by transferring information from richer source domains or systems to sparser targets. DCDCSR combines matrix factorization, sparsity-aware factor construction, and a deep neural network to improve latent-factor mapping and recommendation accuracy.

  • Motivation: CDR and CSR use relatively richer source-domain or source-system information, such as ratings, to improve recommendations in sparser target settings.CDR transfers across domains, whereas CSR transfers across systems.
  • Problem: Existing transfer-based methods map source latent factors to target factors, but accurately modeling their complex cross-domain or cross-system relationship remains difficult.Some approaches replace target factors, while others adjust source factors using common users or items.
  • Approach: DCDCSR is a deep framework that combines matrix-factorization models with a fully connected deep neural network for both CDR and CSR.The matrix-factorization models generate user and item latent factors, while the neural network performs the cross-domain or cross-system mapping.
  • Approach: DCDCSR uses individual users’ and items’ sparsity degrees to combine source and target latent factors into benchmark factors for training.This design is intended to utilize rating data from both domains or systems more effectively.
  • Evaluation: Experiments on three real-world datasets show that DCDCSR outperforms state-of-the-art CDR and CSR approaches in recommendation accuracy.The reported evaluation covers both recommendation settings.

2 Related Work

Prior work addresses cross-domain recommendation through content-based links or transfer-based latent factors and rating patterns. Cross-system recommendation emerged later, with transfer-learning methods that support cross-system item recommendation but may depend on common users or items as bridges.

  • Content-Based Approaches: Content-based CDR links domains using common user or item attributes, merged preferences, social tags, or user-generated texts.These approaches target data sparsity by extracting shared information across domains.
  • Transfer-Based Approaches: Transfer-based approaches use matrix factorization to generate latent factors or rating patterns and transfer them across domains.Prior methods include nonlinear relationships, matrix-based transfer learning, and domain-specific transfer matrices.
  • Cross-System Recommendation: CSR emerged after CDR and applies transfer learning to recommend unrated items across systems.The first CSR work was proposed in Zhao et al. (2013) and later improved in Zhao et al. (2017).
  • Cross-System Recommendation: EMCDR supports both CDR and CSR but uses only common items or common users as a bridge.This limitation distinguishes it from approaches that transfer information without relying exclusively on shared entities.

3 The Proposed DCDCSR Framework

DCDCSR formulates cross-domain and cross-system recommendation as mapping latent factors from source and target data. It combines matrix factorization, sparsity-aware benchmark factors, and a fully connected DNN to generate recommendations.

  • Phase 1: MF Modeling: The framework learns source and target user/item latent factors from rating matrices using matrix factorization.The factorization produces U^s, U^t, V^s, and V^t for subsequent mapping.
  • Phase 3: Recommendation: For CDR, the framework recommends target-domain items to target users using both source and target rating matrices; CSR replaces domains with systems.The CSR formulation is obtained by replacing “domain” with “system” in the CDR definition.
  • Framework overview: DCDCSR uses three phases: MF modeling, DNN mapping, and cross-domain or cross-system recommendation.MF produces latent factors; the DNN maps target factors toward benchmark factors before recommendation.
  • MF model choices: The framework supports MMMF, PMF, and BPR as MF models, introducing rating-oriented and ranking-oriented latent factors into the mapping process.MMMF and PMF focus on rating reconstruction, while BPR emphasizes personalized ranking preservation.
  • Benchmark factor generation: Benchmark factors combine source and target latent factors according to the sparsity degrees of individual users and items.Common entities and different target entities are handled through separate benchmark-factor constructions.
  • Phase 2: DNN Mapping: A fully connected DNN maps normalized target latent factors U^t or V^t to fit benchmark factors U^b or V^b using feedforward and backpropagation.The mapped factors are denormalized before later recommendation steps.

4 Experiments and Analysis

Experiments evaluate DCDCSR across two CDR tasks and two CSR tasks using three real-world datasets, finding that performance generally improves with moderate latent-factor dimensions and that DCDCSR outperforms comparison methods.

  • Experimental setup: Three real-world datasets support two CDR tasks and two CSR tasks evaluated with MAE and RMSE.The tasks cover DoubanMovie→DoubanBook, DoubanMovie→DoubanMusic, Netflix→Douban*Movie, and MovieLens→Douban*Movie.
  • Latent-factor dimension: For latent-factor dimensions K=10 or 20, DCDCSR performance generally improves as K increases; at K=50 improvement is insignificant, while K=100 causes a slight decline.The decline at K=100 is attributed to geometrically increasing DNN parameters while training data remains unchanged.
  • Cross-Domain Recommendation: In CDR, PMF DCDCSR achieves the best MAE, outperforming seven comparison methods by an average of 1.42%, while BPR DCDCSR achieves the best RMSE with an average improvement of 2.6%.The reported improvement ranges are 0.94%–3.57% for MAE and 1.66%–5.41% for RMSE.
  • Cross-System Recommendation: In CSR, BPR DCDCSR outperforms all seven comparison methods by averages of 4.20% in MAE and 4.46% in RMSE.The improvement ranges are 3.00%–9.00% for MAE and 3.43%–9.08% for RMSE.
  • Cross-task comparison: DCDCSR methods deliver larger MAE and RMSE improvements in CSR than CDR because the source systems provide much richer ratings than the target system.The comparison uses MovieLens and Netflix as richer source systems and Douban* as the target system.

5 Conclusions

The paper proposes DCDCSR, a deep framework combining matrix factorization with a fully connected neural network for cross-domain and cross-system recommendation. It guides neural-network training with individual users' and items' sparsity degrees, and reports superior performance on three real-world datasets.

  • Framework: DCDCSR combines matrix factorization models with a fully connected deep neural network to map latent factors across domains or systems.The framework first generates latent factors with matrix factorization and then uses the DNN for mapping.
  • Sparsity guidance: DCDCSR uses individual user and item sparsity degrees from source and target domains or systems to guide DNN training and utilize more rating data.
  • Evaluation: Experiments on three real-world datasets demonstrate superior performance of the proposed model.
Loading 2009.06215v1…