Performance Evaluation of LSTM and GRU Models for Movie Genre Classification Based on Subtitle Dialogs Using Augmented Data and Cross-Validation

This study aims to evaluate and compare the performance of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models in classifying movie genres based on subtitle dialogs. To address data imbalance across genres, data augmentation was applied to create balanced datasets with 500 and 700 samples per genre, in addition to the original dataset. The classification models were built using Word2Vec for word embedding, followed by LSTM and GRU architectures with a single hidden layer and dropout regularization. Model performance was assessed using accuracy and further validated through 5-fold cross-validation. The best test accuracy was achieved with the dataset containing 700 samples per genre, reaching 91% for LSTM and 92% for GRU. Cross-validation showed stable performance with average accuracies of 0.68 for LSTM and 0.67 for GRU. A paired t-test analysis yielded a p-value of 0.341, indicating no statistically significant difference between the two models. These findings suggest that both LSTM and GRU are effective for genre classification based on subtitle dialogs. The use of data augmentation is a key contribution of this study, enabling improved model performance on underrepresented genres. This research supports the development of automated movie recommendation systems that utilize subtitle-based genre prediction.