Application of latent variable models for hidden pattern identification and machine learning prediction improvement in structural engineering

Abstract

Although machine learning models have been widely used in structural engineering, they require predetermined features as input. The input features are usually determined based on observations from previous studies and engineer's experience; therefore, they cannot always guarantee an optimal prediction outcome. To alleviate this critical issue, latent variable models were used in this study. The latent variable models allow for a better machine learning model prediction owing to the incorporation of less correlated latent variables that consider hidden patterns in unobserved data. The superiority of the proposed latent variable modeling approach was explored for two classification problems: (i) failure mode identification of reinforced concrete columns and (ii) seismic damage state identification of bridges. The best latent variable and machine learning model combination for each classification problem was identified based on 10-fold cross-validation. Sparse principal component analysis proved to be the most effective approach for dimensionality reduction. It was found that the use of latent variables can improve the machine learning-based classification accuracy by up to 8 % compared with conventional machine learning models while reducing input dimension by up to ten times significantly improving machine learning training time efficiency.