Development of an AI-Driven Recommendation System for E-Commerce Platforms

Recommendation systems represent one of the most commercially impactful applications of artificial intelligence in e-commerce, directly driving customer engagement, conversion rates, and platform revenue. Existing single-paradigm approaches collaborative filtering, content-based filtering, and matrix factorisation individually suffer from data sparsity, cold-start limitations, and temporal blindness that constrain recommendation quality at scale. This paper presents the development, implementation, and empirical evaluation of a Deep Hybrid Recommendation System (DHRS) for e-commerce platforms, built entirely in Python using a fully reproducible machine learning pipeline. The DHRS integrates Singular Value Decomposition (SVD)-based collaborative filtering, TF-IDF content-based similarity, XGBoost-based temporal feature engineering, Neural Collaborative Filtering (NCF) implemented in PyTorch, and BERT-based semantic NLP processing of product descriptions within a stacked Ridge Regression ensemble. The system was evaluated on a curated dataset of 450,000 user–product interactions from 50,000 users across 5,000 products and 18 categories. Results demonstrate that the proposed DHRS substantially outperforms all ten baseline and comparative models: RMSE = 0.741 (vs. 1.182 popularity baseline; −37.3% improvement), NDCG@10 = 0.627 (vs. 0.304; +106.3%), Precision@10 = 0.648, and catalogue coverage of 78.4%. Ablation studies confirm that each architectural component contributes positively, with collaborative filtering (NDCG contribution: −11.0% when removed) and content-based filtering (−7.8%) providing the largest individual increments. Category-level evaluation reveals strongest performance in Books (NDCG@10 = 0.663, CTR uplift +41.3%) and Electronics (0.647, +34.2%). Cold-start analysis confirms a 31.3% NDCG improvement for users with fewer than five interactions. Seven original figures and six tables document the complete analytical results. The full Python pipeline is documented to support reproducibility and practitioner deployment.