Litigation Outcome Prediction of Differing Site Condition Disputes through Machine Learning Models

Abstract

The construction industry is one of the main sectors of the U.S. economy that has a major effect on the nation’s growth and prosperity. The construction industry’s contribution to the nation’s economy is, however, impeded by the increasing number of disputes that unfold and oftentimes escalate as projects progress. The majority of construction disputes are resolved in courts unless project contracts call for alternate dispute resolution mechanisms. Despite the numerous advantages offered by the litigation process, the extra financial burdens and additional time required by this process makes litigation less desirable in resolving the disputes of a very dynamic construction industry. It is believed that construction litigation could be reduced or even avoided if parties have a realistic understanding of their actual legal position and the likely outcome of their case. Consequently, researchers in the artificial intelligence field have developed tools and methodologies for modeling judicial reasoning and predicting the outcomes of construction litigation cases. Despite the success of some of these systems, they were not on the basis of detailed analyses of legal concepts that govern litigation outcomes. In an attempt to provide a robust legal decision methodology for the construction industry, this paper develops an automated litigation outcome prediction method for differing site condition (DSC) disputes through machine learning (ML) models. To develop the proposed method, this paper compares the performance of three ML techniques, namely: support vector machines (SVMs), naïve Bayes, and rule induction and neural network classifiers (decision trees, boosted decision trees, and the projective adaptive resonance theory). The models were trained and tested using 400 DSC cases filled in the period from 1912 to 2007. Model predictions are on the basis of significant legal factors that govern verdicts in DSC disputes in the construction industry. The third-degree SVM polynomial model performed the best among the nine ML models that were developed, and achieved a prediction precision of 98%.