Usporedba učinaka metoda za smanjenje disbalansa podataka na točnost klasifikacijskih algoritama strojnog učenja

master's thesis

• Title (english): Comparison of the effects of methods for reducing data imbalance on the accuracy of classification algorithms in machine learning
• Author: Kristian Krišto
• Mentor: Mirjana Pejić Bach (mentor)
• Committee member: Mirjana Pejić Bach (predsjednik povjerenstva)
• Committee member: Nikola Vlahović (član povjerenstva)
• Committee member: Katarina Ćurko (član povjerenstva)
• Granter: University of Zagreb; Faculty of Economics and Business; (Department of Informatics); Zagreb
• Defense date and country: 2023-09-29, Croatia
• Scientific / art field, discipline and subdiscipline: SOCIAL SCIENCES; Economics; Business Informatics

Abstract

Veliki faktor kvalitete algoritama strojnog učenja su podaci. Njihova količina i kvaliteta su često bitniji od algoritma koji se koristi. Disbalansirani skupovi podataka nastaju prirodno i imaju izrazito negativan učinak na točnost klasifikacijskih algoritama strojnog učenja koji su na njima trenirani. Kao svojevrsni lijek, postoje različite metode za ublažavanje disbalansa podataka. Teorijski dio rada predstavlja strojno učenje, popularne algoritme, mjere točnosti kao i disbalansirane skupove podataka, njihove izazove te metode za njihovo ublažavanje. U praktičnom dijelu rada se analiziraju efekti tri takve metode na konačnu točnost algoritma. Analiza se provodi koristeći tri skupa podataka i četiri algoritma strojnog učenja kako bi rezultati bili što značajniji. Rad završava zaključkom o tome koju metodu koristiti s obzirom na dani skup podataka i algoritam.

Abstract (english)

Data has a large impact on the quality of machine learning algorithms. It's quantity and quality is often times more important than the choice of algorithm itself. Imbalanced datasets occur naturally and impact the accuracy of classification algorithms in a extremely negative way. To combat this, scientists came up with methods for reducing data imbalance. The theoretical part of this thesis consists of machine learning, popular ML algorithms, accuracy measures as well as imbalanced datasets, their main challenges and the methods used for reducing their imbalance. The practical part focuses on analyzing the effects of three such methods on the final algorithm accuracy. The analysis is done on three different datasets using four different algorithms in an attempt to make the results as general as possible. The thesis ends with a recommendation which method to use with a given dataset and algorithm.

• Keywords: disbalansirani skupovi podataka; strojno učenje; klasifikacija
• Keywords (english): imbalanced data; machine learning; classification
• Language: croatian
• URN:NBN: urn:nbn:hr:148:243747
• Study programme: Title: Business Economics; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: magistar/magistra ekonomije (magistar/magistra ekonomije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Access restricted to students and staff of home institution
• Created on: 2023-10-03 09:57:43