Developing a predictive model concerning insufficiently collected data, and the interaction between variables and parameters – A case study in engineering

Abstract This article presents a practical approach to creating a predictive model that incorporates both known and unknown attributes of the input data. In the conventional procedure, input data are treated as predictors, allowing the response to be obtained. Results of a published paper are us ed; that is an equation relating the probability of failure P f and the coefficient of variation of the soil strength. By manipulating these independent variables and parameters, converting them into non-linear terms, and introducing random noise, a new dataset with various responses and predictors is created. Some values of both predictors and the response in the dataset are deliberately hidden or deleted to simulate missing data. Normalization is required to convert all data to values between 0 and 1. The ‘K-nearest neighbors’ algorithm (kNN) is applied to impute some specified missing data. Results indicate a fairly satisfactory effectiveness. Finally, a dataset without missingness is ready to use some machine learning techniques to construct a predictive model that concerns the interaction between factors and noises in the values of variables and parameters. This study illustrates a simple technique for non-parametric data-driven methods, which could be effectively applied to more complex practical problems in the real world.