Research on a Prediction Model and Importance Algorithm Based on KPCA LightGBM Machine Learning

This paper aims to propose a joint prediction model and factor contribution analysis algorithm based on KPCA-LightGBM. Firstly, kernel principal component analysis (KPCA) is used to reduce the dimension of nonlinear data. Secondly, light gradient boosting machine (LightGBM) is used for data regression and prediction. Finally, MSE and $R^{2}$ were used as model accuracy tests on the predictions of the 31 time series data. The research results show that KPCA can better grasp the characteristics of nonlinear data, such as this paper simplifies 20-dimensional data to 5-dimensional, which can reflect the original data up to more than 95%, and the prediction accuracy of the joint KPCA-LightGBM model is higher compared with the traditional machine learning model, up to 94%.