Application of machine learning methods to classify construction works titles for data normalization and predictive modeling

Abstract

Exceeding the planned deadlines for the implementation of construction projects remains one of the most pressing problems reducing the efficiency of the Russian construction industry. The key obstacle to using historical data and machine learning methods is the extreme fragmentation and denormalization of the source data, in particular, significant variability and non-standardization of the names of construction works. The authors propose a methodology for the automatic unification of variability in the names of construction works using machine learning. This unification is considered as a critical step in building systems for forecasting deadlines within the framework of the state strategy to reduce construction cycles by 30% by 2030. A comprehensive methodology is proposed: data preprocessing, text vectorization and automatic unification. The use of two approaches is studied: clustering (K-Means, DBSCAN) to identify groups of similar works and classification (including Logistic regression, SVM, Random Forest, XGBoost, LSTM) to assign descriptions to unified classes. A comparative analysis of the efficiency of algorithms by the metrics accuracy/recall/F1 was conducted.  High efficiency of machine learning methods (especially XGBoost, Random Forest, LSTM) for unification was proven, which allows to form a structured array of historical data. A comprehensive solution for generating recommendations for optimizing calendar plans based on the analysis of similar historical objects using ML was proposed. A structured array of historical data was formed, including information on the names of works and their durations, which will then be used to implement solutions that allow to obtain reasonable forecasts of deadlines, adjust the planned duration of construction work and minimize the risks of failure, increasing the efficiency of project management