Athens University Economics and Business (AUEB)
Category-based recommender systems: enhancing classical methods using Large Language Models
Abstract
dc:descriptionRecommender systems are deeply embedded in our everyday reality, present in most of the apps and platforms we interact with and are essential operations for global enterprises. But how can smaller providers also incorporate RS into their operations? This research aims to provide guidance for mid-scale providers on implementing RS across different scenarios using a cost-aware approach. The study consists of two parts — a session-aware case and a session-based case — to cover a range of operational situations. For both scenarios, recommendation strategies are suggested based on actual operational conditions, extending the range of methods from basic and conventional to AI-enhanced, as no single method fits all contexts. The session-aware case applies association rules, clustering, and widely used memory-based approaches including collaborative filtering, content-based filtering, and their hybrid combination, to real fashion retail data. The session-based case applies methods such as association rules, Item-KNN, neural networks and an AI-enhanced method that uses LLM-generated descriptions on data from local supermarkets. Findings present that hybrid methods consistently outperform individual methods, and that category-level recommendation is more reliable and indicative than item-level recommendation when sparsity is present, and the item catalog is large and renewed very often. From an algorithmic perspective, descriptors and metadata for items can be game changers in the recommendation process, but they are often limited or unavailable because generating them is demanding. To address this gap, an LLM-enhanced model is proposed that incorporates embeddings generated with minimal human intervention and zero-shot prompting, outperforming all tested benchmarks in the session-based scenario. This study proposes an adaptive framework that guides the choice of recommendation strategy based on data availability and operational conditions, rather than algorithmic complexity alone.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Skiada, Marianna
- Σκιαδά, Μαριάννα
Subjects
dc:subject × 21- Συστήματα προτάσεων
- Συστήματα υπόδειξης
- Πάροχοι μεσαίου μεγέθους
- Μεγάλα γλωσσικά μοντέλα (LLMs)
- Υβριδικά μοντέλα
- Προτάσεις Κατηγοριών
- Αραιότητα δεδομένων
- Recommender systems
- mid-scale providers
- Large Language Models (LLMs)
- Hybrid recommendation methods
- Category recommendations
- Sparsity
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Επιστήμη ηλεκτρονικών υπολογιστών και Πληροφορική, άλλοι τομείς
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Computer and Information sciences, miscellaneous
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62230