Enhancing E-Commerce Recommendations: A Model Proposal Using Advanced Machine Learning

Abstract

E-commerce platforms need intelligent assistants that can understand natural language,identify the right products and provide concise answers without generating false information. This thesis presents a practical system that combines semantic retrieval with a controlled response layer to keep product recommendations aligned with the actual catalog. Product titles, tags and descriptions are encoded using a sentence-level transformer model and then compressed with Principal Component Analysis (PCA) to enable faster nearest-neighbor searches on affordable hardware. The retrieved results are filtered by price, stock and category while an optional reranker balances relevance with business priorities. A lightweight text generator formats short two-line suggestions based only on retrieved products with guardrails to prevent off-catalog or fabricated content. The proposed research delivers a complete, configuration-driven pipeline that covers data curation, embedding, indexing, retrieval and response generation while ensuring reproducibility and transparency. It compares the performance of this approach against keyword-based search and uncompressed embeddings, examining the influence of dimensionality reduction and recommendation length. The evaluation measures ranking quality, catalog coverage, response speed, computational cost and qualitative usefulness across user segments. Findings indicate that PCA maintains retrieval accuracy while significantly improving efficiency and that the controlled response layer enhances clarity without compromising factual accuracy. The study concludes by outlining limitations such as cold-start problems and data drift and suggests future improvements through learning-to-rank methods, personalized ranking strategies and controlled real-world testing to create a scalable, reliable and human-centered recommendation assistant for modern e- commerce platforms.