Sparsity-Aware Movie Recommendation through Dynamic User Profiling and Ensemble Machine Learning

Abstract

Aim: This study aimed to improve movie recommendation accuracy under sparse rating conditions by developing an ensemble system that uses dynamically weighted user profiles to generate personalised recommendations. Design: An experimental quantitative design was adopted. The ensemble recommendation system integrated Case Evaluation using K-means (CE-KM), Case Evaluation using Weighted Association Rule Mining (CE-WARM), and an Adaptive Naive Bayes Probabilistic technique (ADNB). Methodology: The framework comprised three stages: dataset preprocessing and user-profile construction, pattern recognition from users’ movie histories, and the generation and ranking of candidate recommendations. Nine weighted profile attributes represented user interactions and content preferences. The component methods were examined in five experiments using different K values. Accuracy and F-measure were used for evaluation, and the predictions from the three methods were combined to produce the top-ranked movies for an active user. Results: Within the reported experiments, CE-KM achieved its highest accuracy of 0.9323 at K = 70, followed by ADNB which achieved 0.8197 at K = 80 then, CE-WARM which achieved 0.7980 at K = 60. The methods generated different candidate lists, indicating that the profiling, association, and probabilistic components provided complementary outputs. The final ensemble ranking selected movies from the combined predictions. Conclusion: The framework demonstrates a structured approach to combining dynamic user profiling with multiple machine-learning techniques. However, its comparative performance under controlled sparsity conditions requires further evaluation.