Neighborhood component analysis-based feature selection in machine learning to predict tendon ultimate stress of unbonded prestressed concrete beams

Precisely estimating tendon ultimate stress in prestressed concrete beam members while accounting for the impact of various structural parameters holds practical significance in Structural Engineering. Existing approaches in design codes often rely on simplified empirical formulas, frequently falling in accurately representing outcomes across diverse conditions. This paper proposes a novel application of ensemble learning to predict tendon ultimate stress, leveraging relevant parameters identified through Neighborhood Component Analysis (NCA). NCA is strategically utilized for feature selection, enhancing prediction performance and reducing computational cost. An innovative integration of Gradient Boosted Regression Trees (GBRT) with Bayesian optimization for hyper-parameter tuning is introduced, trained, and validated using a robust database of 251 experimental results from simply supported prestressed concrete beams, encompassing a wide range of conditions and structural configurations. The study employs Individual Conditional Expectation (ICE) analysis to elucidate the correlation between input variables and predictions, providing deeper insights into parameter influence. Additionally, the GBRT model is benchmarked against state-of-the-art machine learning algorithms, including Artificial Neural Networks (ANN), Support Vector Machine (SVM), Decision Trees (DT), and Random Forest (RF), demonstrating a good prediction accuracy with R2 of 0.9330. Detailed correlation analysis and benchmarking results highlight the robustness and reliability of the proposed model, underscoring its potential to improve the design of prestressed concrete beam members and offering a significant improvement over traditional empirical approaches.