Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Comparison of traditional machine learning algorithms with large language models for developing personalized recommender systems for enhancing passenger experience on flights

Mähl, Joshua LU (2024) DABN01 20241
Department of Economics
Department of Statistics
Abstract
This thesis compares traditional machine learning algorithms and large language models (LLMs) in developing personalized recommender systems to enhance passenger experiences on flights. Traditional methods like collaborative filtering have been widely used but face challenges such as data sparsity and cold-start problems. LLMs, with their advanced natural language processing capabilities, offer promising solutions to these issues. The study uses a synthetic dataset simulating unary passenger interactions with in-flight entertainment and services, including features like movie preferences, meal choices, and beverage selections. Traditional models, including cosine similarity, singular value decomposition and Bayesian personalized rankings,... (More)
This thesis compares traditional machine learning algorithms and large language models (LLMs) in developing personalized recommender systems to enhance passenger experiences on flights. Traditional methods like collaborative filtering have been widely used but face challenges such as data sparsity and cold-start problems. LLMs, with their advanced natural language processing capabilities, offer promising solutions to these issues. The study uses a synthetic dataset simulating unary passenger interactions with in-flight entertainment and services, including features like movie preferences, meal choices, and beverage selections. Traditional models, including cosine similarity, singular value decomposition and Bayesian personalized rankings, were benchmarked against a recommendation system developed using the GPT-3.5 Turbo model. Evaluation metrics such as accuracy, precision, recall, F1-score, diversity, and coverage were used to assess performance.
Traditional models showed high precision but struggled with recall, leading to lower coverage and diversity. In contrast, the LLM-based system demonstrated superior recall and diversity. This indicates the LLM's effectiveness in identifying a broader range of relevant items, even with sparse data. However, the LLM approach also had trade-offs, including a higher rate of false positives and still significant popularity bias. Despite these challenges, the LLM outperformed traditional methods in providing diverse and accurate recommendations. (Less)
Please use this url to cite or link to this publication:
author
Mähl, Joshua LU
supervisor
organization
course
DABN01 20241
year
type
H1 - Master's Degree (One Year)
subject
keywords
recommendation systems, large language models, collaborative filtering
language
English
id
9161962
date added to LUP
2024-09-24 08:35:47
date last changed
2024-09-24 08:35:47
@misc{9161962,
  abstract     = {{This thesis compares traditional machine learning algorithms and large language models (LLMs) in developing personalized recommender systems to enhance passenger experiences on flights. Traditional methods like collaborative filtering have been widely used but face challenges such as data sparsity and cold-start problems. LLMs, with their advanced natural language processing capabilities, offer promising solutions to these issues. The study uses a synthetic dataset simulating unary passenger interactions with in-flight entertainment and services, including features like movie preferences, meal choices, and beverage selections. Traditional models, including cosine similarity, singular value decomposition and Bayesian personalized rankings, were benchmarked against a recommendation system developed using the GPT-3.5 Turbo model. Evaluation metrics such as accuracy, precision, recall, F1-score, diversity, and coverage were used to assess performance.
Traditional models showed high precision but struggled with recall, leading to lower coverage and diversity. In contrast, the LLM-based system demonstrated superior recall and diversity. This indicates the LLM's effectiveness in identifying a broader range of relevant items, even with sparse data. However, the LLM approach also had trade-offs, including a higher rate of false positives and still significant popularity bias. Despite these challenges, the LLM outperformed traditional methods in providing diverse and accurate recommendations.}},
  author       = {{Mähl, Joshua}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Comparison of traditional machine learning algorithms with large language models for developing personalized recommender systems for enhancing passenger experience on flights}},
  year         = {{2024}},
}