Title:RODIAN: Robustified Median

Abstract:We propose a robust method for averaging numbers contaminated by a large proportion of outliers. Our method, dubbed RODIAN, is inspired by the key idea of MINPRAN [1]: We assume that the outliers are uniformly distributed within the range of the data and we search for the region that is least likely to contain outliers only. The median of the data within this region is then taken as RODIAN. Our approach can accurately estimate the true mean of data with more than 50% outliers and runs in time $O(n\log n)$. Unlike other robust techniques, it is completely deterministic and does not rely on a known inlier error bound. Our extensive evaluation shows that RODIAN is much more robust than the median and the least-median-of-squares. This result also holds in the case of non-uniform outlier distributions.