A Concept Lattice-Based Kernel for SVM Text Classification

Standard Support Vector Machines (SVM) text classification relies on bag-of-words kernel to express the similarity between documents. We show that a document lattice can be used to define a valid kernel function that takes into account the relations between different terms. Such a kernel is based on the notion of conceptual proximity between pairs of terms, as encoded in the document lattice. We describe a method to perform SVM text classification with concept lattice-based kernel, which consists of text pre-processing, feature selection, lattice construction, computation of pairwise term similarity and kernel matrix, and SVM classification in the transformed feature space. We tested the accuracy of the proposed method on the 20NewsGroup database: the results show an improvement over the standard SVM when very little training data are available.