Adaptive Sparse Bayesian Regression with Variational Inference for Parameter Estimation

A relevance vector machine (RVM) is a sparse Bayesian modeling tool for regression analysis. Since it can estimate complex relationships among variables and provide sparse models, it has been known as an efficient tool. On the other hand, the accuracy of RVM models strongly depends on the selection of their kernel parameters. This article presents a kernel parameter estimation method based on variational inference theories. This approach is quite adaptive, which enables RVM models to capture nonlinearity and local structure automatically. We applied the proposed method to artificial and real datasets. The results showed that the proposed method can achieve more accurate regression than other RVMs.