Novel Machine Learning Models Based Uncertainty Estimation and Sequential Predictions for Blockchain Networks.

Abstract

Cryptocurrencies based-blockchains – decentralised banks of public ledgers of transactions and pseudonymous identities – lure criminals to incognito behind an alphanumeric address to con- duct illicit activities over the network. Consequently, this double-edged sword technology urges the necessity of analysing blockchain data to detect illicit activities. In the existing literature, visual analytics have been widely used to gain useful insights from large-scale blockchain data via graph network analysis and visual analytics tools. However, a straightforward visualisation is not very effective with the increasing complexity of the blockchain network. On the other hand, a machine learning approach is capable of dealing with the massive amount of data generated by the public blockchain. Machine learning techniques have provided promising results in many big data applications e.g. social media, IoT, and recently blockchain. Nev- ertheless, the public blockchain is subject to unseen events that the machine learning model might not be aware of. In this work, machine learning models are developed in a novel way and uncertainty quantification methods are exploited to detect illicit activities in the public blockchain. Various supervised learning models are thoroughly explored using two datasets derived from Bitcoin and Ethereum blockchains. The unexpected events of blockchain are ad- dressed by computing uncertainty estimates besides the machine learning model’s predictions. Consequently, a novel uncertainty estimation method is proposed that reveals an effective per- formance in comparison to existing methods. In this context, uncertainty estimation reflects the model’s uncertain predictions about a given input. Subsequently, two distinct frame- works are presented that serve different purposes using blockchain data. The first framework is viewed as an end-to-end prototype of the temporal-GNN classification model – temporal graph neural network – based on an active learning process to reduce the laborious labelling process of blockchain data. In particular, the active learning approach utilises the predicted uncertainties to query the most informative data points where the active learning framework is performed and evaluated using a variety of acquisition functions. The other framework presents a novel model based on sequential predictions. This model refers to RecGNN – a recurrent graph convolutional network – that requires the predictions of the antecedent nodes in the Bitcoin transaction graph as input features. As a result, this project shows the effectiveness of using tree-based classifiers in classi- fying data derived from the public blockchain. This allows the detection of illicit activities (e.g. illicit transactions, users, accounts) that operates over the blockchain. It also highlights the competence of models based on graph learning algorithms throughout this project. The active learning frameworks using the temporal-GNN model have revealed promising results. Moreover, the highest performance provided by the RecGNN model is discussed to classify illicit Bitcoin transactions with an accuracy of 98.09% and f1-score of 91.75%. The experi- mental results are evaluated using classification metrics and other statistical measurements. The limitations, challenges and possible future directions are also demonstrated. Keywords: Machine Learning, Supervised Learning, Graph Neural Network, Temporal Graph Neural Network, Uncertainty Estimation, Bayesian Methods, Active Learning, Sequential Prediction, Blockchain.