A high-performance extreme gradient boosting outlier detection framework for integrating the outputs of diverse anomaly detectors for detecting mineralization-related geochemical anomalies

Abstract

In geochemical exploration, the geochemical anomalies identified in the same area by different unsupervised anomaly detection models are often quite divergent. How to combine these divergent geochemical anomalies into reliable mineral prospecting targets is a problem worth studying. In this regard, the extreme gradient boosting outlier detection (XGBOD) framework was adopted to integrate the anomaly scores produced by diverse unsupervised anomaly detection models to construct a high-performance semi-supervised anomaly detection ensemble for detecting mineralization-related geochemical anomalies. In the XGBOD framework, various unsupervised anomaly detection models are built and used to transform input variables into the transformed outlier scores (TOSs), and the important TOSs are then selected and added into the original input data to train the extreme gradient boosting (XGBoost) model, and a high-performance semi-supervised XGBoost model is established finally for detecting mineralization-related geochemical anomalies. The superiority of the XGBOD framework was demonstrated by a case study implemented in the Baishan area (Jilin, China). The K-nearest neighbor, local outlier factor, histogram-based outlier score, one-class support vector machine and isolation forest were used to transform element concentrations to TOSs, and the TOSs were used as the input data of the XGBoost model together with the original input element concentration data. The XGBoost model was finally established to detect mineralization-related geochemical anomalies. The results show that the semi-supervised XGBoost model performs significantly better than the five unsupervised anomaly detection models. Therefore, the XGBOD framework is a viable tool for combining diverse anomaly scores produced by various anomaly detectors to build a high-performance semi-supervised ensemble for detecting mineralization-related geochemical anomalies.