Feature weights in contractor safety performance assessment: Comparative study of expert-driven and analytics-based approaches

Abstract

Current expert-based approaches to determining the weights of different safety management elements during contractor safety performance are time-consuming and potentially biased.Hence, this paper evaluates analytics-based approaches, i.e., supervised learning, cluster-then-predict and two-level variable weighting K-Means (TWKM) (an extension of the traditional K-Means clustering algorithm), against the Delphi method. In collaboration with an infrastructure developer, a dataset of 461 data points and 12 features describing subcontractors' inherent risks and safety assurance performance were collected. This paper showed that supervised learning improves recall by 21 % when compared with the Delphi method. This paper also highlights that changes in input features' distributions (or covariate shifts) across construction stages and projects can reduce the recall of the supervised learning model from 93 % to 50 %. Key academic and practical contributions include the analytics-based approaches to develop weights for measuring contractors' safety performance, and strategies to manage the impact of covariate shifts on accuracy of feature weights.