A geometric framework for outlier detection in high‐dimensional data

Outlier or anomaly detection is an important task in data analysis. We discuss the problem from a geometrical perspective and provide framework which exploits metric structure of set. Our approach rests on manifold assumption, that is, observed, nominally high-dimensional lie much lower dimensional this intrinsic can be inferred with learning methods. show exploiting significantly improves outlying observations high data. also suggest novel, mathematically precise widely applicable distinction between distributional structural outliers based geometry topology clarifies conceptual ambiguities prevalent throughout literature. experiments focus functional as one class structured data, but we propose completely general include image graph applications. results outlier non-tabular detected visualized using methods quantified standard scoring applied to embedding vectors. This article categorized under: Technologies > Structure Discovery Clustering Fundamental Concepts Data Knowledge Visualization