<![CDATA[This paper examines the use of a Restricted Boltzmann Machine (RBM) to provide personalized product recommendations on a niche coffee e-commerce platform facing cold-start conditions. We train RBM variants on a binary transaction matrix derived from 100 simulated user transactions and evaluate four hidden-unit configurations (3, 5, 10, 15) using 5-fold cross-validation. Models were trained with Contrastive Divergence (CD-1) and assessed primarily by Mean Squared Error (MSE) for reconstruction fidelity, complemented by ranking metrics (Precision@3, NDCG@3). The 10-hidden-unit configuration achieved the best balance of reconstruction and ranking performance, with an average test MSE ? 0.0454, outperforming popular-item (MSE: 0.0802) and random (MSE: 0.0760) baselines. While the RBM demonstrates strong capability in modeling latent user preferences under sparse data, ranking metrics expose limitations when predicting exact top-N items in extremely sparse cases. The study highlights practical implications for early-stage niche marketplaces and suggests integrating content signals or hybridization to further improve top-N recommendation quality.]]>
Copyrights © 2025
