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