Kernels on Lists and Sets over Relational Algebra: An Application to Classification of Protein Fingerprints

摘要

In this paper we propose a new class of kernels defined over extended relational algebra structures. The "extension" was recently proposed in [1] and it overcomes one of the main limitation of the standard relational algebra, i.e. difficulties in modeling lists. These new kernels belong to the class of R-Convolution kernels in the sense that the computation of the similarity between two complex objects is based on the similarities of objects' parts computed by means of sub-kernels. The complex objects (relational instances in our case) are tuples and sets and/or lists of relational instances for which elementary kernels and kernels on sets and lists are applied. The performance of this class of kernels together with the Support Vector Machines (SVM) algorithm is evaluated on the problem of classification of protein fingerprints and by combining different data representations we were able to improve the best accuracy reported so far in the literature.