Mining social and cryptocurrency networks

Abstract

A general problem with graph data is that it cannot be fed to classical machine learning methods in a straightforward way. Algorithms like logistic regression or decision trees only work well with tabular data. Due to the irregular size of node neighborhoods, raw network data cannot be considered tabular. Recently, several static node embedding models were proposed to learn a vector space representation of network nodes for downstream machine learning tasks. Unfortunately, static graph mining models do not perform well in data-intensive tasks where interactions between network participants are constantly arriving over time. Fitting batch algorithms for large graph snapshots could cause a significant time-delay in the prediction. That is why online graph learning techniques are much preferred in these scenarios. In this thesis, I analyze user interactions in social and cryptocurrency networks with user-related metadata that can be used as ground truth for most of the addressed graph mining tasks. Specifically, we intend to answer the following questions: (1) What are the main advantages of online graph mining techniques over batch models for large-scale social networks and how to best compare their performance? In our research, we focus on graph centrality and node embedding techniques and we propose three online algorithms for these domains. (2) How to mine cryptocurrency networks with novel network science tools to answer open questions in the domain of cryptoeconomics and privacy? By collecting various new Twitter and cryptocurrency network data sets, we were among the first to deploy and analyze node embedding models in several network applications such as vaccine skepticism detection or Ethereum address deanonymization.