Zgodnje odkrivanje prevar pri menjavi kriptovalut s strojnim učenjem : magistrsko delo

Tehovnik, Lea (author), Todorovski, Ljupčo (mentor), Demšar, Jaka (comentor)

Z nastankom novih tehnologij, kot so kriptovalute, se pojavijo tudi novi načini prevar. Soočimo se s problematiko bančnih prevar na kripto borzah. Prevare želimo identificirati in jih ustaviti, preden je prepozno. Primarno za to uporabimo metode za nadzorovano strojno učenje, kjer se osredotočimo na drevesne metode, natančneje na naključne gozdove. Napovedujemo redek pojav, zato imamo opravka z zelo neuravnoteženo porazdelitvijo ciljne spremenljivke, kar rešujemo z uporabo metod pod- in nad-vzorčenja učnih podatkov. Končni model smo zgradili z metodo naključnih gozdov, kategorične spremenljivke pa smo pretvorili v numerične, saj smo pokazali, da to pripomore k izboljšavi rezultatov. Rezultati na testnih podatkih kažejo, da je končni model uporaben, saj identificira skoraj vse prevarante. Spodbuden rezultat lahko uporabimo za nadaljnji razvoj modela in implementacijo v sistem borze.

• Language: Slovenian
• Keywords: strojno učenje, naključni gozdovi, napovedovanje redkih dogodkov, odkrivanje prevar, kriptovalute
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2020
• PID: 20.500.12556/RUL-121701
• UDC: 004.4
• COBISS.SI-ID: 33726467
• Publication date in RUL: 23.10.2020

TEHOVNIK, Lea, 2020, Zgodnje odkrivanje prevar pri menjavi kriptovalut s strojnim učenjem : magistrsko delo [online]. Master’s thesis. [Accessed 26 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=121701

Secondary language

• Language: English
• Title: Early-stage detection of crypto-trading frauds with machine learning

Abstract

With the emergence of new technologies, such as cryptocurrencies, new types of fraud are born. We focus on bank fraud common on crypto exchanges. We want to classify and distinguish fraudulent cases, so they can be prevented in advance. We use machine learning algorithms for classification based on decision trees, more specifically, random forests, i.e., ensembles of decision trees. Predicting the rare event of fraud leads to a classification problem with unbalanced distribution of the target variable, which we address with using methods for over- and under-sampling of learning data. We build the final model with random forest method alongside with transformation of categorical variables to numerical, which was proved to improve the results. Test results show that the end model is very useful, as it identifies almost all fraudulent cases. Encouraging results can be used for further development of the model and its implementation into the exchanges' system.

• Keywords: machine learning, random forests, predicting rare events, fraud detection, cryptocurrencies