An Unsupervised Heterograph Model for Identifying Operational Risk Events in Financial Services

Operational risk in financial services is one of three main risks confronted by commercial banks. This risk can cause commercial banks to suffer huge losses every year. Effective protection against such risk has received widespread attention from academic researchers. These researchers usually discuss operational risk management at the macrostrategic level rather than at the micro-operational level. Nevertheless, we have found that preventing potential operational risk events in advance can significantly reduce the losses of commercial banks. In this process, identifying abnormal operational events plays a critical role. Unlike previous studies, this paper regards operating subjects, actions, and objects in an operational event as a whole and associates them using a heterograph. To address the problem of missing labels, we introduce an unsupervised heterogeneous graph model to identify operational risk events. In particular, this model comprises two modules: a heterograph autoencoder and a GMM estimation network. The heterograph autoencoder is responsible for learning the low-dimensional node embeddings containing the neighbourhood and semantic information of the heterograph data. The GMM estimation network aims to evaluate the likelihood for each node using these node embeddings and recognize the nodes with a high likelihood as anomalies according to a threshold. In addition, to improve the performance of identifying operational risk events, we build models from both temporal and spatial perspectives. A series of experiments on our unique dataset demonstrate that our model outperforms state-of-the-art models in identifying operational risk events. Finally, we open-source the code of our model and our unique data to GitHub.