This paper proposes a deep learning-based movie recommendation system that integrates a transformer architecture with KMeans clustering to enhance the accuracy and diversity of recommendations. The system addresses challenges such as data sparsity, scalability, and the cold-start problem by leveraging user behavior sequences, movie embeddings, and KMeans clustering for genre-based recommendations. The model uses a transformer layer with positional encoding and multi-head attention mechanisms to capture temporal dependencies in user preferences. It then integrates KMeans clustering on movie genre embeddings to group similar movies and enhance recommendation diversity. The system was evaluated on two MovieLens datasets (100 K and 1 M), achieving significant improvements in RMSE, MAE, precision, recall, and F1 scores compared to baseline models. The results demonstrate that the proposed system effectively mitigates cold-start and scalability issues while providing more accurate and diverse recommendations. The model's performance was further validated through experiments on different latent dimensions, showing that increasing the embedding dimension improves recommendation quality. The integration of KMeans clustering with the transformer model enhances the system's ability to provide personalized and diverse recommendations, making it effective in the context of abundant data.This paper proposes a deep learning-based movie recommendation system that integrates a transformer architecture with KMeans clustering to enhance the accuracy and diversity of recommendations. The system addresses challenges such as data sparsity, scalability, and the cold-start problem by leveraging user behavior sequences, movie embeddings, and KMeans clustering for genre-based recommendations. The model uses a transformer layer with positional encoding and multi-head attention mechanisms to capture temporal dependencies in user preferences. It then integrates KMeans clustering on movie genre embeddings to group similar movies and enhance recommendation diversity. The system was evaluated on two MovieLens datasets (100 K and 1 M), achieving significant improvements in RMSE, MAE, precision, recall, and F1 scores compared to baseline models. The results demonstrate that the proposed system effectively mitigates cold-start and scalability issues while providing more accurate and diverse recommendations. The model's performance was further validated through experiments on different latent dimensions, showing that increasing the embedding dimension improves recommendation quality. The integration of KMeans clustering with the transformer model enhances the system's ability to provide personalized and diverse recommendations, making it effective in the context of abundant data.