Construction Dispute Potentials: Mechanism versus Empiricism in Artificial Neural Networks

The booming development of neural network algorithms has shifted the research focus in the field of construction project management from causal investigation to statistical approximation and hence from mechanistic models to empirical models. This paper took construction dispute avoidance as an example and enabled the best efforts to establish paired mechanistic and empirical models to investigate if the pursuit of a mechanistic understanding of construction disputes should be continued. A Bayesian belief network and multilayer perceptron were used for mechanistic and empirical simulations, respectively. A list of critical dispute factors was first identified from the literature and shortlisted by Pearson’s chi-square tests and Pearson product-moment correlational coefficient tests. The structure of the Bayesian belief network was constructed with logical deduction assisted by a further literature review and Delphi surveys. A structured questionnaire survey was conducted to collect quantitative data for factor shortlisting and model quantification. It was revealed that, being assisted with machine learning techniques, both mechanistic and empirical models achieved an accuracy rate of over 95% under ideal conditions. However, Bayesian belief network models predicted better with fewer constraints due to their advantages in reflecting the formation mechanism of construction disputes, while multilayer perceptron models were more constrained by the inconvenience of sourcing high-quality data as model input. This paper demonstrated that it is still necessary to investigate the formation mechanism of construction disputes further for more efficient avoidance strategies. During the investigation of model construction and comparison, this paper also reflected on the interpretation of statistical threshold and proposed that an arbitrary single cut-off point for statistical tests could potentially eliminate factors that should have been included.