Anomaly detection in the probability simplex under different geometries

Abstract An open problem in data science is that of anomaly detection. Anomalies are instances do not maintain a certain property present the remaining observations dataset. Several detection algorithms exist, since process itself ill-posed mainly because criteria separates common or expected vectors from anomalies unique. In most extreme case, labelled and algorithm has to identify anomalous, assign degree each vector. The majority make any assumptions about properties feature space which embedded, may affect results when those spaces properties. For instance, compositional such as normalized histograms, can be embedded probability simplex, constitute particularly relevant case. this contribution, we address detecting relying on concepts Information Geometry, by focusing our efforts distance functions commonly applied context. We report series experiments conclude specific distance-based relies Geometry-related instead Euclidean distance, performance significantly improved.