Personalized Movie Recommendation Based on User Preferences Using Optimized Sequential Transformer Model

This study explores the application of a Transformer based model to the “MovieLens 1M” and “MovieLens 10M” datasets provided by GroupLens, which includes one million and ten movie ratings. The Transformer model, known for its proficiency in handling sequence-to-sequence tasks, employs multi-head self-attention mechanisms and position-wise feed-forward networks. Our implementation, developed in PyTorch, benefits from its capacity for automatic differentiation and GPU acceleration, which significantly enhances the training efficiency. The model underwent rigorous hyperparameter tuning and optimization to adapt optimally to the characteristics of the dataset. Performance evaluation using the Normalized Discounted Cumulative Gain (NDCG) metric demonstrates that the Transformer model substantially outperforms a conventional popular recommendation system across various top-K rankings. Specifically, for the MovieLens 1M dataset, the Transformer model achieved NDCG scores of 0.0455, 0.0805, 0.0995, and 0.1273 at top 1, 3, 5, and 10, respectively. Comparatively, the popular recommendation system scored 0.0023, 0.0044, 0.0061, and 0.0091 at these same rankings. This trend of superior performance by the Transformer model is also consistent across the MovieLens 10M dataset, further confirming its effectiveness in ranking highly relevant items at the top of the list. Also, to address quality and utility of recommendations, MAP, Precision, Recall, F1 Score, Coverage and Serendipity metrics evaluated. These results underscore the effectiveness of Transformer models in extracting and leveraging complex patterns in sequence data, demonstrating a clear superiority over traditional methods in recommendation systems. This study highlights the potential of deep learning architectures to revolutionize fields reliant on understanding user preferences and sequential data interactions.