Selection of Subsets of Ordered Features in Machine Learning

摘要

The new approach of relevant feature selection in machine learning is proposed for the case of ordered features. Feature selection and regularization of decision rule are combined in a single procedure. The selection of features is realized by introducing weight coefficients, characterizing degree of relevance of respective feature. A priori information about feature ordering is taken into account in the form of quadratic penalty or in the form of absolute value penalty on the difference of weight coefficients of neighboring features. Study of a penalty function in the form of absolute value shows computational complexity of such formulation. The effective method of solution is proposed. The brief survey of author's early papers, the mathematical frameworks, and experimental results are provided.