Gated Recurrent Unit-Based Deep Learning Framework for Personalized and Sequential Smart TV Content Recommendation

With the rapid growth of smart TVs, the development of efficient recommendation systems has become essential to enhance the user experience. Most existing recommendation systems rely on collaborative filtering and matrix factorization techniques, which are limited in capturing temporal dependencies and sequential user behaviour. To overcome these issues, this research uses Gated Recurrent Units with an attention mechanism for smart TV content, leveraging sequential data to understand temporal patterns in user viewing sequences. This utilizes cross dataset validation on movielens100k and 1 M dataset for ensuring scalability. The proposed recommender statistical significance pipeline is designed with Item embeddings, stacked GRU layers, and a lightweight attention mechanism to minimize computational overhead on resource-constrained edge devices. Model performance has been evaluated using fivefold cross-validation and computational Flops with identical architectures. The proposed model has produced a normalized Root Mean Square Error (RMSE) of 0.25 with MAE validation confirmed using a paired t-test at a 0.05 significance level, demonstrating stable performance across different datasets. In contrast, the graph-based model with an autoencoder has 0.8, the Deep Belief network algorithm with Monarch Butterfly Optimisation has 0.9, and KNN and the restricted Boltzmann machine learning algorithm have 0.8. The proposed model shows significant improvement, achieving normalized RMSE and better generalization across k-fold validations for real time recommendation scenarios.