Identification of illicit blockchain transactions using hyperparameters auto-tuning

The rapid increase in interest in cryptocurrencies has elevated the importance of developing trusted methods for combating the field's numerous illegal activities. The advantage of finance activity based on a blockchain for investigation is that the framework provides many raw data points for free, which experts can query and analyze using machine learning algorithms to build automated frameworks for preventing illicit behavior. Previous research has demonstrated this possibility, paving the way for cryptocurrency fraud to be carefully avoided. The chapter proposes improvements in classifying Bitcoin transactions over previous research by utilizing recent automated machine learning algorithms for auto-tuning the hyperparameters of the models. The chapter presents a quantitative description of several experiments conducted with the goal of investigating the Elliptic dataset. First, the setup for the machine learning experiments is described, and then the results are presented. The main findings concern the LightGBM classifier, which achieved an F1 score of 82% during the evaluation phase, whereas the top F1 score reported in literature is around 79%. The improvements are due to the model's hyperparameters being auto-tuned utilizing a parameters search algorithm.