Design and Verification of a Genre-Based Movie Recommendation System Using Movie Attributes and Random Forest

Abstract

The objective of this study is to solve two critical issues of streaming platform movie recommendation systems: the "cold start" problem (ineffectiveness for new users/new movies) and the challenge of balancing popularity and quality in genre-specific recommendations. These are vital because over 80% of streaming user activities depend on such systems, and flawed recommendations lower user retention (Netflix Technology Blog, 2021).

Methods formed the core of this study: First, the public MovieLens 100K dataset (100,000 ratings, 943 users, 1,682 movies) was used, with preprocessing including cleaning (removing movies with <10 ratings), encoding genre strings via one-hot encoding, and standardizing continuous features (Z-Score). Second, a genre-based system relying solely on movie attributes (rating count v, average rating R, genres) was built. Two models—linear regression (parametric, linear assumption) and random forest (non-parametric, ensemble of decision trees)—were compared, with performance evaluated using R² and RMSE. Results showed random forest outperformed linear regression (R²=0.9917 vs. 0.9682, RMSE=0.123 vs. 0.215), solving cold start issues and generating user-matching recommendations.