Anomaly Detection based on Association Rules and Conformal Prediction


We propose a novel technique based on a combination of association rule learning and conformal prediction in its label-dependent form. The non-conformity score is not based directly on a classification algorithm, but aggregates information of many association rules that can be extracted from the data, and exceptions from them. As an application, we use customer data from UKB. There are multiple fields indicating if a customer is an Industrial Corporation (INC) or Small/Medium-sized Enterprise (SME). We will consider these as labels. Often these labels are incorrect or inconsistent across the SAP system. The aim is to use machine learning to identify inconsistencies and potential errors.

Proceedings of Machine Learning Research 105:1-2, 2019