Journal Article

Hybrid K-Nearest Neighbors Models with Metaheuristic Optimization for Predicting Undrained Shear Strength

Cao, Ning, Xu, Genqi, Ma, Jing, Zhang, Lijuan, Yan, Xin-E

2025

Overview

Around the world, soft soils can be found in many areas close to seas and rivers. These areas play an indispensable and crucial role in the development of government plans, especially in the population growth sector. Due to maintaining a weak shear power and vast settlement under the buildings, soft soils are considered problematic soil. The significant risks associated with building structures and infrastructures in soft soil are high, requiring engineers' extreme attention. It depends on undrained shear strength (USS) that the foundation of structures can bear in soft soil, and this factor vigorously controls the selection of soil improvement techniques. In recent years, there have been enhancements and extensions in the methodologies employed for estimating soil characteristics, including USS. These methods are divided into three main sections: Laboratory Testing, Field Testing, and Correlation with Other Soil Parameters. In recent research, data science techniques have created more reliable and accurate models for predicting USS. This study aims to apply the K-Nearest Neighbors (KNN) classifying method for predicting USS. Mountain Gazelle Optimizer (MGO) and Coronavirus Herd Immunity Optimizer (CHIO) appeal for developing hybrid models with KNN and facilitating accuracy enhancement. The dataset which utilized in this study contains four input variables including liquid limit (LL), plastic limit (PL), and sleeve friction (SF), overburden weight (OBW). Comparative analysis across all data phases reveals that the KNCH model, optimized using the CHIO, achieved superior predictive performance with the highest coefficient of determination (R² = 0.993), and the lowest values in root mean square error (RMSE = 85.19), mean squared error (MSE = 7256.15), normalized RMSE (NRMSE = 0.470), and scatter index (SI = 0.065). In contrast, the KNN model without optimization reported R² = 0.971, RMSE = 168.17, and SI = 0.132, while the KNMG model—optimized using the MGO—resulted in R² = 0.983, RMSE = 128.15, and SI = 0.101.

• Publisher: Slovenian Society Informatika / Slovensko drustvo Informatika
• Subject: Algorithms / Civil engineering / Construction / Coronaviruses / Data science / Deep learning / Friction / Herd immunity / Heuristic methods / Human error / K-nearest neighbors algorithm / Liquid limits / Machine learning / Optimization / Optimization techniques / Overburden / Performance evaluation / Plastic limit / Population growth / Root-mean-square errors / Shear strength / Shear tests / Soft soils / Soil improvement