Towards automatic feature construction for supervised classification

Abstract

We suggest an approach to automate variable construction for supervised learning, especially in the multi-relational setting. Domain knowledge is specified by describing the structure of data by the means of variables, tables and links across tables, and choosing construction rules. The space of variables that can be constructed is virtually infinite, which raises both combinatorial and over-fitting problems. We introduce a prior distribution over all the constructed variables, as well as an effective algorithm to draw samples of constructed variables from this distribution. Experiments show that the approach is robust and efficient. © 2014 Springer-Verlag.