by Agata Skorupka
The study examines different graph-based methods of detecting anomalous activities on digital markets, proposing the most efficient way to increase market actors’ protection and reduce information asymmetry. Anomalies are defined below as both bots and fraudulent users (who can be both bots and real people). Methods are compared against each other, and state-of-the-art results from the literature and a new algorithm is proposed. The goal is to find an efficient method suitable for threat detection, both in terms of predictive performance and computational efficiency. It should scale well and remain robust on the advancements of the newest technologies. The article utilized three publicly accessible graph-based datasets: one describing the Twitter social network (TwiBot-20) and two describing Bitcoin cryptocurrency markets (Bitcoin OTC and Bitcoin Alpha). In the former, an anomaly is defined as a bot, as opposed to a human user, whereas in the latter, an anomaly is a user who conducted a fraudulent transaction, which may (but does not have to) imply being a bot. The study proves that graph-based data is a better-performing predictor than text data. It compares different graph algorithms to extract feature sets for anomaly detection models. It states that methods based on nodes’ statistics result in better model performance than state-of-the-art graph embeddings. They also yield a significant improvement in computational efficiency. This often means reducing the time by hours or enabling modeling on significantly larger graphs (usually not feasible in the case of embeddings). On that basis, the article proposes its own graph-based statistics algorithm. Furthermore, using embeddings requires two engineering choices: the type of embedding and its dimension. The research examines whether there are types of graph embeddings and dimensions that perform significantly better than others. The solution turned out to be dataset-specific and needed to be tailored on a case-by-case basis, adding even more engineering overhead to using embeddings (building a leaderboard of grid of embedding instances, where each of them takes hours to be generated). This, again, speaks in favor of the proposed algorithm based on nodes’ statistics. The research proposes its own efficient algorithm, which makes this engineering overhead redundant.