Dynamical Representation Learning for Ethereum Transaction Network via Non-negative Adaptive Latent Factorization of Tensors

Abstract

As a common cryptocurrency platform, Ethereum involves massive accounts and numerous real-time transactions. Moreover, as the involved accounts increase drastically, it is impossible to have transactions among all accounts at one time slot, which results in a high-dimensional and incomplete (HDI) dynamic transaction network. In spite of its HDI nature, such HDI dynamic transaction network contains much useful knowledge regarding involved accounts' behavior patterns like potential transaction links. To extract such knowledge from an HDI dynamic transaction network, this paper proposes a Non-negative Adaptive Latent Factorization of Tensors (NAL) model with two interesting ideas: a) adopting an HDI tensor to describe an HDI dynamic transaction network and building a non-negative learning objective based on the principle of data density-oriented, and b) implementing model hyper-parameter self-adaptive via using a particle swarm optimization (PSO) algorithm in the training process. Empirical studies on three real Ethereum transaction networks show that compared with state-of-the-art methods, the proposed NAL model achieves superior performance in terms of accuracy and computational efficiency in predicting potential transaction links.