Convolution Kernels for Outliers Detection in Relational Data

Abstract

There is a growing interest to kernel-based methods in Data Mining. The application of these methods for real-world data, stored in databases, leads to the problem of designing kernels for complex structured data. Since many Data Mining systems use relational databases, the important task is to design kernels for relational data. In this paper we show that for relational data the structure of single data instance in the input space can be described by nested relation schemes. For such data we propose the method for constructing kernels, which is based on convolution kernels framework developed by Haussler. For demonstration we construct the simple convolution Gaussian kernel and apply it, using k-nearest neighbor algorithm, for outliers detection problem in the sample relational database.