Can machine learning approach classify making-do waste cases in construction sites?

Abstract

Purpose Making-do waste represents losses caused by improvisation on construction sites. Such losses are commonly described in written texts. This loose description makes it difficult and laborious to read and classify them. To better understand these losses, this article investigates the viability of using machine learning (ML) to classify making-do cases, highlighting its potential to streamline the analysis process traditionally conducted by researchers and practitioners. The study began by examining the typology of input data used for classifying waste cases, followed by a survey of potential algorithmic models capable of performing this classification.

Design/methodology/approach This research uses a database containing 6,399 losses over 9 vertical building projects of Brazilian construction companies. Each loss was manually labeled into a missing prerequisite, category and impact. Using the input data, which is written text in Brazilian Portuguese, we trained several ML models. The dataset was split in a training and a test part to understand the effectiveness of the algorithms.

Findings The study found that, on average, the ML models were able to correctly classify above 90% of the cases. With respect to the techniques, the combination of neural networks and k-nearest neighbors (KNN) through stacking significantly improved the accuracy and efficiency of the methods.

Research limitations/implications The study relied on a dataset written in Brazilian Portuguese for training and validation, which may limit the generalizability of the results. Future research could explore additional datasets and further optimize the models to increase their robustness.

Originality/value This research contributes to the literature by demonstrating the effective application of ML in waste classification, reducing the labor-intensive nature of manual analysis and providing a robust methodology for future studies in this domain.