B1695
Title: Anomaly edge detection in network data using conformal prediction
Authors: Rui Luo - City University of Hong Kong (Hong Kong) [presenting]
Abstract: Conformal prediction is a user-friendly paradigm for generating set-valued predictions for machine learning models that are valid in a distribution-free sense. It demonstrates how conformal prediction can be used to detect anomalous edges in a network by exploiting edge exchangeability as a criterion for distinguishing anomalous edges from normal ones. To quantify the difference between a given edge and existing normal edges in the graph, the variational inference is used to approximate the inverse transaction posterior probability, which serves as the non-conformity score. An anomaly detector is then presented based on the conformal prediction that has a guaranteed upper bound for the false positive rate. Through numerical experiments, the proposed algorithm is shown to achieve comparable performance in detecting anomalous transactions in a blockchain network when compared to baseline methods. The results demonstrate the effectiveness of using conformal prediction and variational inference for detecting anomalous transactions in blockchain.
