Adaptive cluster-based outlier detection

The analysis of data is typically accompanied by concern as to the correctness of recorded data points; some of the points might be contaminated, thereby distorting the result of the analysis. This paper proposes a novel cluster-based and distribution-independent method for outlier detection. Based on Monte Carlo simulations, the new method is tested with different data distributions and compared with the method of standardised residuals (also known as the z-score). It is shown that the cluster-based approach identifies outliers more reliably, even for a normal data distribution, and the advantages are discussed in detail.