Recommendation systems are an integral and indispensable part of every digital platform, as they can suggest content or items to users based on their respective needs. Collaborative filtering is a technique often used in various studies, which produces recommendations by analyzing similarities between users and items based on their behavior. Although often used, traditional collaborative filtering techniques still face the main challenge of sparsity. Sparsity problems occur when the data in the system is sparse, meaning that only a portion of users provide feedback on some items, resulting in inaccurate recommendations generated by the system. To overcome this problem, we developed a Hybrid Collaborative Filtering model based on Matrix Factorization and Gradient Boosting (HCF-MFGB), a new hybrid approach. Our proposed model integrates SVD++, the XGBoost ensemble learning algorithm, and utilizes user demographic data and meta items. We utilize information, both explicitly and implicitly, to learn user preference patterns using SVD++. The XGBoost algorithm is used to create hundreds of decision trees incrementally, thereby improving model accuracy. Meanwhile, user demographic and meta-item data are clustered using the K-Means Clustering algorithm to capture similarities in user and item characteristics. This combination is designed to improve rating prediction accuracy by reducing reliance on minimal explicit rating data, while addressing sparsity issues in movie recommendation systems. The results of experiments on the MovieLens 100K, MovieLens 1M, and CiaoDVD datasets show significant improvements, outperforming various other baseline models in terms of RMSE and MAE. On the MovieLens 100K dataset, the HCF-MFGB model obtained an RMSE value of 0.853 and an MAE value of 0.674. On the MovieLens 1M dataset, the HCF-MFGB model obtained an RMSE value of 0.763 and an MAE value of 0.61. On the CiaoDCD dataset, the HCF-MFGB model achieved an RMSE value of 0.718 and an MAE value of 0.495. These results confirm a significant improvement in movie recommendation accuracy with the proposed approach.
- Article type
- Year
Open Access
Article
Issue
Open Access
Article
Issue
An imbalanced dataset often challenges machine learning, particularly classification methods. Underrepresented minority classes can result in biased and inaccurate models. The Synthetic Minority Over-Sampling Technique (SMOTE) was developed to address the problem of imbalanced data. Over time, several weaknesses of the SMOTE method have been identified in generating synthetic minority class data, such as overlapping, noise, and small disjuncts. However, these studies generally focus on only one of SMOTE’s weaknesses: noise or overlapping. Therefore, this study addresses both issues simultaneously by tackling noise and overlapping in SMOTE-generated data. This study proposes a combined approach of filtering, clustering, and distance modification to reduce noise and overlapping produced by SMOTE. Filtering removes minority class data (noise) located in majority class regions, with the k-nn method applied for filtering. The use of Noise Reduction (NR), which removes data that is considered noise before applying SMOTE, has a positive impact in overcoming data imbalance. Clustering establishes decision boundaries by partitioning data into clusters, allowing SMOTE with modified distance metrics to generate minority class data within each cluster. This SMOTE clustering and distance modification approach aims to minimize overlap in synthetic minority data that could introduce noise. The proposed method is called “NR-Clustering SMOTE,” which has several stages in balancing data: (1) filtering by removing minority classes close to majority classes (data noise) using the k-nn method; (2) clustering data using K-means aims to establish decision boundaries by partitioning data into several clusters; (3) applying SMOTE oversampling with Manhattan distance within each cluster. Test results indicate that the proposed NR-Clustering SMOTE method achieves the best performance across all evaluation metrics for classification methods such as Random Forest, SVM, and Naїve Bayes, compared to the original data and traditional SMOTE. The proposed method (NR-Clustering SMOTE) improves accuracy by 15.34% on the Pima dataset and 20.96% on the Haberman dataset compared to SMOTE-LOF. Compared to Radius-SMOTE, this method increases accuracy by 3.16% on the Pima dataset and 13.24% on the Haberman dataset. Meanwhile, compared to RN-SMOTE, the accuracy improvement reaches 15.56% on the Pima dataset and 19.84% on the Haberman dataset. This research result implies that the proposed method experiences consistent performance improvement compared to traditional SMOTE and its latest variants, such as SMOTE-LOF, Radius-SMOTE, and RN-SMOTE, in solving imbalanced health data with class binaries.
京公网安备11010802044758号