The effectiveness of data pre-processing methods on the performance of machine learning techniques using RF, SVR, Cubist and SGB: a study on undrained shear strength prediction

In the field of data engineering in machine learning (ML), a crucial component is the process of scaling, normalization, and standardization. This process involves transforming data to make it more compatible with modeling techniques. In particular, this transformation is essential to ensure the suitability of the data for subsequent analysis. Despite the application of many conventional and relatively new approaches to ML, there remains a conspicuous lack of research, particularly in the geotechnical discipline. In this study, ML-based prediction models (i.e., RF, SVR, Cubist, and SGB) were developed to estimate the undrained shear strength (UDSS) of cohesive soil from the perspective of a wide range of data-scaling and transformation methods. Therefore, this work presents a novel ML framework based on data engineering approaches and the Cubist regression method to predict the UDSS of cohesive soil. A dataset including six different features and one target variable were used for building prediction models. The performance of ML models was examined considering the impact of the data pre-processing issue. For that purpose, data scaling and transformation methods, namely Range, Z-Score, Log Transformation, Box-Cox, and Yeo-Johnson, were used to generate the models. The results were then systematically compared using different sampling ratios to understand how model performance varies as various data scaling/transformation methods and ML algorithms were combined. It was observed that data transformation or data sampling methods had considerable or limited effects on the UDSS model performance depending on the algorithm type and the sampling ratio. Compared to RF, SVR, and SGB models, Cubist models provided higher performance metrics after applying the data pre-processing steps. The Box-Cox transformed Cubist model yielded the best prediction performance among the other models with an R2 of 0.87 for the 90% training set. Also, the UDSS prediction model generally yielded the best performance metrics when it was used with the transformed-based models (i.e., Box-Cox, Log, and Yeo-Johnson) than that of scaled-based (i.e., Range and Z-Score) models. The results show that the Cubist model has a higher potential for UDSS prediction, and data pre-processing methods have impacts on the predictive capacity of the evaluated regression models.