Predicting Construction Workforce Demand Using a Combination of Feature Selection and Multivariate Deep-Learning Seq2seq Models
Milad Ashtab, Boong Yeol Ryoo
Abstract
Construction companies struggle with their hiring plans and react to economic shifts afterward, resulting in unnecessary layoffs and overhirings. A model that forecasts future construction hiring can help adjust hiring levels based on upcoming projections. This research proposes a framework for predicting the future sequence (upcoming 12 months) of hiring values instead of specific months, based on historical data between 1993 and 2022 for hiring and economic explanatory variables. Explanatory variables are categorized into the local, neighboring states, and national levels. Feature selection methods were used to filter out the initial data set to reduce data dimensionality—the output of each method trained by the recurrent neural network (RNN). Seq2seq models were evaluated based on their mean absolute error (MAE). The results of the best-performing model indicate that the multivariate seq2seq model can capture general trends and disruptions due to economic recession and natural disasters more accurately than the univariate statistical models, even though there was no feature inside the data set regarding hurricanes.