Scholarlink

<p>Data sparsity is one of the primary drawbacks of Single-Domain Recommender Systems, which results in suboptimal user–item pairings and erroneous recommendations. To overcome this limitation, in this paper, we introduce the Cross-Domain Recommendation System (CDRS), which facilitates knowledge transfer across different domains without requiring shared users or items. The method follows Orthogonal Non-Negative Matrix Tri-Factorization (ONNMTF) to group users and items in the Auxiliary Domain (AD) and constructs a pattern matrix using the model to record latent rating structures. This matrix is then taken to the Target Domain (TD) to establish user–item correlations and to predict missing ratings to reduce sparsity. The predicted ratings in the TD are further refined using Pearson Correlation Coefficient (PCC) similarity and K-Nearest Neighbors (K-NN) to create final predictions. The system’s accuracy is evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) as the evaluation metrics on two datasets namely MovieLens-10&#xa0;k and Amazon-Food. Performance of the proposed system is compared with nine top-level approaches (including SVD-, SVD + + -, and RMGM-based frameworks plus three deep learning-based models) which show that the accuracy of our proposed method is superior. The proposed CDRS gives a minimum of 0.5670 MAE and RMSE of 0.8983, outperforming the best baseline by 4.51% for MAE, while still achieving computational economy over 10–100 iterations. Collectively, the proposed ONNMTF-oriented cross-domain framework is beneficial as it effectively migrates user–item interaction patterns across heterogeneous domains, alleviates data sparsity, and further improves the accuracy and reliability of recommendation systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-domain recommender system for non-overlapping domains using clustering-based pattern transfer mechanism

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Antara Bhattacharya,
  • Nishat A. Ansari,
  • Ashish Tiwari

摘要

Data sparsity is one of the primary drawbacks of Single-Domain Recommender Systems, which results in suboptimal user–item pairings and erroneous recommendations. To overcome this limitation, in this paper, we introduce the Cross-Domain Recommendation System (CDRS), which facilitates knowledge transfer across different domains without requiring shared users or items. The method follows Orthogonal Non-Negative Matrix Tri-Factorization (ONNMTF) to group users and items in the Auxiliary Domain (AD) and constructs a pattern matrix using the model to record latent rating structures. This matrix is then taken to the Target Domain (TD) to establish user–item correlations and to predict missing ratings to reduce sparsity. The predicted ratings in the TD are further refined using Pearson Correlation Coefficient (PCC) similarity and K-Nearest Neighbors (K-NN) to create final predictions. The system’s accuracy is evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) as the evaluation metrics on two datasets namely MovieLens-10 k and Amazon-Food. Performance of the proposed system is compared with nine top-level approaches (including SVD-, SVD + + -, and RMGM-based frameworks plus three deep learning-based models) which show that the accuracy of our proposed method is superior. The proposed CDRS gives a minimum of 0.5670 MAE and RMSE of 0.8983, outperforming the best baseline by 4.51% for MAE, while still achieving computational economy over 10–100 iterations. Collectively, the proposed ONNMTF-oriented cross-domain framework is beneficial as it effectively migrates user–item interaction patterns across heterogeneous domains, alleviates data sparsity, and further improves the accuracy and reliability of recommendation systems.