Implementasi Variational Autoencoder dan Regularized Singular Value Decomposition pada Sistem Rekomendasi Film

Abstract

The abundance of available films presents a challenge for viewers in finding content that matches their preferences. This study proposes a hybrid ensemble movie recommendation system combining a Variational Autoencoder (VAE) with KL annealing and Regularized Singular Value Decomposition (RSVD) using a weighted average approach on the MovieLens 100K dataset. Evaluation is conducted using 10-fold cross-validation, final testing on a hold-out test set, and statistical significance testing via paired t-test. Results show that the hybrid ensemble with weights alpha = 0.25 for VAE and beta = 0.75 for RSVD achieves an RMSE of 0.9307 ± 0.0043 and MAE of 0.7338 ± 0.0034 on cross-validation, and RMSE of 0.9365 and MAE of 0.7401 on the test set, consistently outperforming standalone RSVD (RMSE 0.9474) and standalone VAE (RMSE 1.0926). All performance differences are confirmed statistically significant by paired t-test with p-values well below 0.05.