Abstract: Recent advances in recommendation systems have demonstrated significant potential through deep learning approaches yet challenges remain in computational efficiency and prediction accuracy. This research presents a novel framework that integrates deep learning compression techniques with statistical variance analysis to enhance movie recommendation performance while reducing computational overhead. The proposed system leverages multi-modal data from MovieLens and IMDB datasets, implementing vector quantization and embedding compression to achieve optimal memory utilization. Our methodology incorporates standard deviation analysis to evaluate recommendation consistency and employs quantization-aware training for model optimization. Experimental validation using MovieLens 25M and IMDB datasets demonstrates superior performance compared to baseline collaborative filtering methods. The system achieves a 15.3% reduction in Root Mean Squared Error (RMSE) while maintaining 79.2% compression ratio through INT4 quantization. Statistical analysis reveals improved recommendation consistency with standard deviation values of 0.89 for highly-rated content compared to 1.67 for polarizing content. The framework addresses critical gaps in computational efficiency and recommendation accuracy particularly in large-scale deployment scenarios. Results indicate significant improvements in both prediction quality and system efficiency with 68% reduction in memory requirements and 45% faster inference time. This research contributes to the advancement of efficient recommendation systems by demonstrating the effectiveness of combining compression techniques with statistical analysis for enhanced user experience and system scalability.
Keywords: Movie Recommendation, Deep Learning Compression, Vector Quantization, Standard Deviation Analysis, Movie Lens Dataset, IMDB Integration, Collaborative Filtering, Embedding Compression
|
DOI:
10.17148/IJARCCE.2025.14591
