Pristupi strojnog učenje u uvjetima neuravnoteženih podataka

master's thesis

• Author: Maja Kovačić
• Mentor: Tomislav Šmuc (mentor)
• Committee member: Tomislav Šmuc (predsjednik povjerenstva)
• Committee member: Pavle Pandžić (član povjerenstva)
• Committee member: Filip Najman (član povjerenstva)
• Committee member: Goran Muić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2019-12-04, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

Pretpostavka u problemima klasifikacije jest da je broj primjera među različitim klasama podjednak, što nije slučaj u problemima iz stvarnog svijeta. Taj problem je poznat kao učenje u uvjetima neuravnoteženih podataka. Učenje u takvim uvjetima nailazi na mnoge poteškoće, poput pristranosti većinskoj klasi dok je manjinska klasa od interesa. Tu su važne metode za balansiranje podataka kako bi smanjili efekt iskrivljenosti distribucije klasa. Ovdje promatramo metode prekomjernog i pod- uzorkovanja. Metode prekomjernog uzorkovanja mogu bolje inducirati klasifikatore koji su precizniji nego oni inducirani metodom pod-uzorkovanja. Te metode kombiniramo s ansamblima, kao bagging i boosting. Modifikacija ansambla uključuje tehnike za balansiranje prije učenja klasifikatora i ovdje su promatrane razne kombinacije s ciljem pronalaska najboljeg klasifikatora. U nastavku, u ovom radu provodimo eksperimentalnu evaluaciju koja uključuje pet metoda za rad s nebalansiranim podacima na deset skupova podataka s UCI repozitorija. Dalje se bavimo statističkom validacijom rezultata, te proučavanjem da li algoritmi odstupaju u performansama u odnosu na ostale.

Abstract (english)

Assumption in classification problems is that number of examples in different classes is approximately equal, which is not case in real world problems. The problem is known as learning in conditions of imbalanced data. Learning in such conditions encounters many difficulties, such as bias to the majority class while minority class is interested one. There are important methods for balancing data to minimize skewness of class distribution. Here are observed over-sampling and under-sampling methods. Over-sampling methods can better induce classifier which are more accurate than those induced by under-sampling methods. These methods are combined with ensembles, such as bagging and boosting. Ensemble modification includes balancing techniques before learning the classifiers and here are given different combinations with best classifier purpose. In addition, in this work we perform a board experimental evaluation involving five methods to deal with class imbalanced problems in ten UCI datasets. We are further concerned on statistical validation of results and by examining whether algorithms differ in performance compared to others.

• Keywords: neuravnoteženi podaci; prekomjerno uzorkovanje; pod-uzorkovanje; ansambli; bagging; boosting; statistička validacija
• Keywords (english): imbalanced data; over-sampling; under-sampling; ensembles; bagging; boosting; statistical validation
• Language: croatian
• URN:NBN: urn:nbn:hr:217:720971
• Study programme: Title: Mathematical Statistics; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra matematike (magistar/magistra matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2020-07-03 13:08:51