PREDICTIVE INFORMATION CRITERIA FOR ROBUST RELEVANCE VECTOR REGRESSION MODELS

Access this Article

Search this Article

Abstract

The relevance vector regression (RVR) is a Bayesian nonlinear regression procedure whose model is expressed in terms of kernel functions, like the support vector regression (SVR). In order to overcome the sensitivity to outliers of the RVR, the robust relevance vector regression (RRVR) procedures have been proposed. A crucial issue in the model building process of the RRVR is the choice of kernel parameters. The selection of these parameters can be viewed as a model selection and evaluation problem. In this paper, we derive a model selection criterion for the Bayesian predictive distribution of the RRVR models. Monte Carlo experiments and real data analysis show that our model selection criterion performs well in various situations.

Journal

  • Bulletin of informatics and cybernetics

    Bulletin of informatics and cybernetics (50), 65-80, 2018-12

    統計科学研究会

Codes

  • NII Article ID (NAID)
    120006620470
  • NII NACSIS-CAT ID (NCID)
    AA10634475
  • Text Lang
    ENG
  • Article Type
    departmental bulletin paper
  • ISSN
    0286-522X
  • Data Source
    IR 
Page Top