Benchmarking Unsupervised Outlier Detection with Realistic Synthetic Data

Benchmarking unsupervised outlier detection is difficult. Outliers are rare, and existing benchmark data contains outliers with various unknown characteristics. Fully synthetic usually consists of regular instances clear characteristics thus allows for a more meaningful evaluation methods in principle. Nonetheless, there have only been few attempts to include benchmarks detection. This might be due the imprecise notion or difficulty arrive at good coverage different domains data. In this work, we propose generic process generation datasets such benchmarking. The core idea reconstruct from real-world while generating so that they exhibit insightful We describe benchmarking detection, as sketched far. then three instantiations generate specific characteristics, like local outliers. To validate our process, perform state-of-the-art carry out experiments study quality reconstructed way. Next showcasing workflow, confirms usefulness proposed process. particular, yields close ones real Summing up, new practical