Detection of Outliers in Geochemical Data Using Ensembles of Subsets of Variables

Geochemical data used in geological interpretation of mine deposits and identification of geological domains often contain outliers. Undertaking statistically sound and robust decision-making about outliers (such as deciding whether observations under consideration belong to a given domain) can be a challenging task. Traditional statistical procedures are often poorly suited to the noisy, intrinsically multivariate and high-dimensional nature of geochemical data. We present herein a novel approach for detecting outliers robustly in large multi-dimensional geochemical data. The approach incorporates a feature selection method that automatically seeks the best subset of chemical ratios that, together with the original chemical variables, best represent the inherent characteristics of the data. The proposed approach robustly distinguishes outliers even at high contamination levels. Experimental results demonstrating the advantages of the proposed feature selection algorithm over previous methods used in outlier detection are shown using data from an iron ore deposit located in the Brockman Iron Formation of Hamersley Province, Western Australia.