Data Dimensionality Reduction by Introducing Structural Equation Modeling to Machine Learning Problems

Structural Equation Modeling (SEM) is a statistical analysis technique for fitting networks of constructs to data. The strength of extracting latent factors and estimating casual relations among them makes it accessible to analyze the structure of social problems, such as those around the economy, healthcare, and ecology. The present paper introduces SEM to the machine learning field for reducing the number of data dimensions. Too large dimensionality of the input data increases the computational load to build machine learning models. The proposed method utilizes SEM to sift variables and sort out relations between factors. The results show that the proposed method is able to identify the least number of data dimensions for machine learning-based prediction models to achieve the same or even higher accuracy as the prediction models trained with the full dimensions does.