IOGOD: An interpretable outlier generation-based outlier detector for categorical databases

With the property of interpretability, applications can provide valuable information on the functioning of machine learning models. Interpretability is important in applications where explanations are a must, as in loan rejections, and it allows feedback to executives and analysts on what a models’ results are based on. However, the best performing outlier detectors are not interpretable, and therefore, their results cannot be taken advantage of effectively, or used for relevant applications. Hence, we propose IOGOD, an interpretable outlier generation-based outlier detector, which uses an outlier generation stage to train a two-class outlier detection to achieve comparable performance to state-of-the-art black-box algorithms. IOGOD uses decision trees within an autoencoder to learn the structure of the genuine data and generate synthetic outlier samples, which are then used to train an interpretable two-class contrast pattern-based classifier. Our experimental results show that IOGOD outperforms other interpretable outlier detectors and performs comparably to black-box outlier detectors in comparisons against 14 outlier detector algorithms on 39 categorical databases taken from the KEEL machine learning repository.