Authors:
Nuno Gomes
1
and
Artur Ferreira
2
;
1
Affiliations:
1
ISEL, Instituto Superior de Engenharia de Lisboa, Instituto Politécnico de Lisboa, Portugal
;
2
Instituto de Telecomunicações, Lisboa, Portugal
Keyword(s):
Bitcoin, Feature Reduction, Feature Selection, Fraud Detection, Supervised Learning.
Abstract:
The use of cryptocurrencies corresponds to a remarkable moment in global financial markets. The nature of cryptocurrency transactions, done between cryptographic addresses, poses many challenges to identify fraudulent activities, since malicious transactions may appear as legitimate. Using data with these transactions, one may learn machine learning models targeted to identify the fraudulent ones. The transaction datasets are typically imbalanced, holding a few illicit examples, which is challenging for machine learning techniques to identify fraudulent transactions. In this paper, we investigate the use of a machine learning pipeline with dimensionality reduction techniques over Bitcoin transaction data. The experimental results show that XGBoost is the best performing method among a large set of competitors. The dimensionality reduction techniques are able to identify adequate subsets suitable for explainability purposes on the classification decision.