Optimizing Recommendation Performance with a Multi-Stage k-Means, DNN, RBM, and k-NN Pipeline
Ayush Yajnik, Vishal Sharma · International Journal of Science and Research (IJSR) · 2025
Recommendation systems help users navigate vast item catalogs, yet traditional collaborative-and content-based filtering suffer from data-sparsity, cold-start issues, and limited ability to model complex user item relationships. To overcome these challenges, we present a unified hybrid pipeline that first partitions the item space with k-means clustering, then employs a deep neural network for feature extraction and cluster selection, refines selections through a Restricted Boltzmann Machine, and finally delivers item suggestions using k-Nearest Neighbors. Experiments on the Kaggle Spotify dataset after z-score normalization and SMOTE-based class-balancing show that our deep network attains an average F1-score of 0.97, RBM refinement boosts within-cluster accuracy, and the final k-NN stage yields superior Precision, Recall, NDCG, and MAP compared with baseline collaborative-filtering and matrix-factorization models. These results demonstrate that orchestrating complementary algorithms in a multi-stage workflow produces robust, scalable, and highly accurate recommendations suitable for real-world deployment.