Modifying the Tailored Clustering Enabled Regionalization (TCER) framework for outlier site detection and inference efficiency

Abstract

The “regional advantage” hypothesizes that inference uncertainty/prediction error of geotechnical or geological properties at a target site (a site contains a group of records measured from different locations/depths using a variety of tests) can be smaller if we use a quasi-regional cluster (includes two or more database sites with geotechnical or geological properties similar to the target site) instead of the entire database. A tailored clustering enabled regionalization (TCER) framework has been proposed to verify this “regional advantage” hypothesis. TCER requires the target site should not be an outlier site relative to the database. However, it remains a challenge on how to detect an outlier site (or data group) from a database. In this paper, we modify the original TCER by introducing a novel outlier site detection step called maximum site similarity (MSS) into the original TCER. The capability of MSS is verified using synthetic and real examples. Additionally, three inference methods [e.g., probabilistic multiple regression (PMR), classical Bayesian model (CBM), and hierarchical Bayesian model (HBM)] are studied for the purpose of determining the optimal inference method for the modified TCER in terms of achieving the minimum inference uncertainty/prediction error with reasonable computational time. It is shown that the modified TCER with CBM outperforms other inference methods for the examples shown in this paper.