03790ntm a22005417i 4500 000725516 CZ-PrVSE 20260201122909.0 m d cr n|||||||||| 260201s2025 xr fsbm 000 0 eng d NEZPRACOVANÝ IMPORT ABA006 cze ABA006 ABA006 rda Melnychenko, Hlib ISIS:174688 dis Vylepšení doporučování v e-commerce pomocí strojového učení a grafových databází eng Enhancing E-Commerce Recommendations with Machine Learning and Graph Databases / Hlib Melnychenko 2025 ?? stran : digital, PDF soubor Vedoucí práce: Jiří Feuerlicht Diplomová práce (Ing.)—Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky, 2026 Obsahuje bibliografii Textový (vysokoškolská kvalifikační práce) Rok obhajoby 2026 This thesis tackles the critical challenge of delivering relevant product recommendations in e-commerce environments where data sparsity and "Cold-Start" users (those with no history) render traditional algorithms ineffective. To solve this, the work proposes and implements a novel Inductive Hybrid Recommendation System. This architecture uniquely combines structural learning with deep semantic understanding, utilizing GraphSAGE—an inductive Graph Neural Network—to model complex user-item relationships, and Sentence-BERT (SBERT) to generate rich semantic embeddings from unstructured product text. These features are synthesized by a supervised XGBoost Ranker, optimized for a Pairwise Ranking objective to ensure high-quality list ordering. Evaluated on a large-scale Amazon Reviews dataset using a strict chronological split to prevent data leakage, the hybrid model demonstrated superior robustness. While a traditional Content-Based Baseline failed completely for new users (0% success), the Hybrid model achieved a 15% Hit Rate for cold-start users, effectively solving the zero-start problem. Simultaneously, it outperformed the baseline for established users with an 84.6% improvement in ranking quality (nDCG). Beyond quantitative success, the system leverages the Neo4j graph database to visualize decision pathways, providing interpretability and trust in the recommendations. The thesis concludes by outlining a deployment roadmap and proposing future enhancements such as Multi-Modal Feature Fusion using CLIP to integrate visual data for even richer discovery. Způsob přístupu: Internet data a analytika pro business [obor dipl. práce] diplomové práce fd132022 czenas master's theses eczenas Recommender systems SBERT Graph neural network Graph SAGE XGBoost Feuerlicht, Jiří pna2008466870 ths Potančok, Martin vse20201071395 opn Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky kn20010709399 dgg https://insis.vse.cz/zp/91857/podrobnosti VŠKP v InSIS https://insis.vse.cz/zp/91857 Hlavní práce https://insis.vse.cz/zp/91857/posudek/vedouci Hodnocení vedoucího https://insis.vse.cz/zp/91857/posudek/oponent/88165 Oponentura https://insis.vse.cz/zp/91857/priloha/32895 Přiloha k práci https://insis.vse.cz/zp/91857/priloha/33240 Přiloha k práci https://insis.vse.cz/zp/91857/podrobnosti dc:identifier NEPOSILAT VSKP vse91857 260121 91857