What question did this study set out to answer?

To develop and evaluate a hybrid collaborative filtering model for enhancing recommendation accuracy based on historical user interactions.

March 13, 2026Open Access

Product recommendation framework with enhanced item-based collaborative filtering

AKArchana Praveen KumarManipal Academy of Higher Education CChinmaiManipal Academy of Higher Education BKB. Priya KamathManipal Academy of Higher Education

Key Points

To develop and evaluate a hybrid collaborative filtering model for enhancing recommendation accuracy based on historical user interactions.
Hybrid collaborative filtering model developed
Compared with user-user collaborative filtering and neural collaborative filtering
Assessed accuracy using Food.com and MovieLens datasets
Evaluated performance using root mean square error (RMSE) metrics
Conducted bias analysis across user demographics
Hybrid model achieved RMSE of 0.5015, a 12.4% improvement over mean-centered CF
Neural Collaborative Filtering yielded lowest RMSE of 0.7138 for MovieLens
Item-item CF, SVD outperformed traditional CF with RMSEs of 0.8019 and 0.9368 respectively
Bias analysis showed consistent variations in errors across user demographics
Hybrid model more effective for sparse data, NCF better for dense contexts

Abstract

Abstract Recommendation systems play a very important role and are therefore essential in the modern digital platforms that aid users in identifying content aligned to their preferences based on historical interactions. The Collaborative Filtering (CF) is adopted widely as it provides relationships between the user and item that are used to generate the recommendations. The proposed work develops and assesses a hybrid CF model and compares it with two baseline methods: User-user CF and Neural CF. The study analyzes the accuracy of the recommendations generated by the hybrid model across different domains through hybridization. Two datasets used in this experiment are Food.com recipe ratings and MovieLens, representing distinct domains. The former dataset is experimented with raw and mean-centered Item-item CF as well as Singular Value Decomposition (SVD) based Matrix Factorization and a weighted hybrid model. The Hybrid model achieved the best performance with an RMSE of 0.5015, improving prediction accuracy by 12.4% over mean-centered CF (0.5725 RMSE). For the MovieLens dataset, Item-Item CF, SVD, and Neural Collaborative Filtering (NCF) were evaluated, with NCF yielding the lowest prediction error (0.7138 RMSE) and outperforming SVD (0.8019) and CF (0.9368). Bias analysis indicated minor but consistent variations in residual error across user demographics. Overall, the Hybrid CF model proved more effective for sparse food-rating data, while NCF better captured complex preference structures in dense movie-rating contexts.

Perguntar à IA

Bookmark

View Full Paper