Scholarlink

The speedy Proliferation of the Internet’s user generator contents led up to the appearance of huge amounts of data, products and services. Collaborative filtering (CF) approaches are magic tools depend on K nearest-neighbors, they have received significant attention and remain the most popular recommender systems due to their simplicity, efficiency and ability to result precise and personalized recommendations for products and serves by estimating user reaction to the items evaluation. The objective of this paper is to present a model that optimized KNN-based recommendation system model which employs memory-based Collaborative filtering (item-based) by fine-tunes and evaluates the performance (accuracy) for two different variant prediction benchmarks. Similarity computations using three distinct similarity measures—Cosine Similarity, Pearson Correlation Coefficient and Mean Squared Difference (MSD)—were conducted, to compare their impact on rating estimation precise in Collaborative filtering (CF) models. The experimental study conducted using popular two benchmark datasets known as Movilens 1M and FilmTrust. The study results indicate that the enhanced model delivers improved performance, when applying the experiment using Movielens 1M dataset, we obtained better performance (accuracy) using KNNWithMeans when computing the similarity using both cosine, person correlation co-efficient and MSD similarity measures, by (10.8%), (10.2%) and (2.6%) respectively in terms of RMSE, and (7.9%), (9.8%) and (2.2%) respectively in terms of MAE. The KNNBasic algorithm slightly outperformed the KNNWithMeans algorithm in performance (accuracy) when applying the experiment on a small dataset (FilmTrust) by (10.8%), (10.2%) and (2.6%) respectively in terms of RMSE, and (7.9%), (9.8%) and (2.2%) respectively in terms of MAE.

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

Fine-Tuning KNN-Based Recommender Systems for Personalized Recommendations: A Study on Three Similarity Measures and Their Performance

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hussam Elbehiery,
  • Mohamed Mahmoud Fouad,
  • Fainan Nagy El-Sisi,
  • Asma Haroun Elsaid

摘要

The speedy Proliferation of the Internet’s user generator contents led up to the appearance of huge amounts of data, products and services. Collaborative filtering (CF) approaches are magic tools depend on K nearest-neighbors, they have received significant attention and remain the most popular recommender systems due to their simplicity, efficiency and ability to result precise and personalized recommendations for products and serves by estimating user reaction to the items evaluation. The objective of this paper is to present a model that optimized KNN-based recommendation system model which employs memory-based Collaborative filtering (item-based) by fine-tunes and evaluates the performance (accuracy) for two different variant prediction benchmarks. Similarity computations using three distinct similarity measures—Cosine Similarity, Pearson Correlation Coefficient and Mean Squared Difference (MSD)—were conducted, to compare their impact on rating estimation precise in Collaborative filtering (CF) models. The experimental study conducted using popular two benchmark datasets known as Movilens 1M and FilmTrust. The study results indicate that the enhanced model delivers improved performance, when applying the experiment using Movielens 1M dataset, we obtained better performance (accuracy) using KNNWithMeans when computing the similarity using both cosine, person correlation co-efficient and MSD similarity measures, by (10.8%), (10.2%) and (2.6%) respectively in terms of RMSE, and (7.9%), (9.8%) and (2.2%) respectively in terms of MAE. The KNNBasic algorithm slightly outperformed the KNNWithMeans algorithm in performance (accuracy) when applying the experiment on a small dataset (FilmTrust) by (10.8%), (10.2%) and (2.6%) respectively in terms of RMSE, and (7.9%), (9.8%) and (2.2%) respectively in terms of MAE.