Usporedba algoritama temeljenih na paradigmi gradient boostinga: Primjena na stvarne medicinske i fizikalne skupove podataka

master's thesis

• Title (english): Comparison of algorithms based on the gradient boosting paradigm: Application to real medical and physical datasets
• Author: Alan Brkičić
• Mentor: Davor Horvatić (mentor)
• Committee member: Davor Horvatić (predsjednik povjerenstva)
• Committee member: Katarina Jeličić (član povjerenstva)
• Committee member: Maro Cvitan (član povjerenstva)
• Committee member: Maja Planinić (član povjerenstva)
• Committee member: Tomislav Bokulić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Physics); Zagreb
• Defense date and country: 2024-09-27, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Physics

Abstract

U ovom radu se proučavaju razni pristupi strojnom učenju s naglaskom na algoritme bazirane na paradigmi gradient boostinga, ulazi se u njihovu matematičku pozadinu i u dubine funkcioniranja algoritama. Uspoređuje se efikasnost gradient boostinga, XGBoosta, LightGBM-a, CatBoosta i slučajne šume na dva skupa podataka. Skup podataka Higgsovog bozona nastoji identificirati Higgsov bozon unutar raznih signala. Skup podataka za emboliju nastoji predvidjeti hoće li pacijent preživjeti do svoje sljedeće kontrole. Rezultati navedenih algoritama se mjere njihovom točnošću, opozivom, preciznošću i ROC-AUC vrijednosti. Mjeri se i značajnost značajki pripadnih skupova podataka pomoću Shapleyjevih vrijednosti.

Abstract (english)

This thesis examines various approaches to machine learning, with a focus on algorithms based on the gradient boosting paradigm. It delves into their mathematical foundations and the intricacies of their inner workings. The efficiency of gradient boosting, XGBoost, LightGBM, CatBoost, and random forests is compared using two datasets: the Higgs boson dataset, which aims to identify the Higgs boson among various signals, and the embolism dataset, which seeks to predict patient outcomes based on disease severity and other comorbidities. The performance of these algorithms is evaluated using accuracy, recall, precision, and ROC-AUC values. Additionally, the signficance of the features in each dataset is assessed using Shapley values.

• Keywords: strojno učenje; Gradient boosting; XGBoost
• Keywords (english): machine learning; Gradient boosting; XGBoost
• Language: croatian
• URN:NBN: urn:nbn:hr:217:145560
• Study programme: Title: The Integrated University Undergraduate and Graduate Programme in Physics and Computer Science Education; specializations in: Physics and informatics; Course: Physics and informatics; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: sveučilišni magistar edukacije fizike i informatike (sveučilišni magistar edukacije fizike i informatike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2025-01-31 12:55:10