Materializing and Querying Learned Knowledge

Abstract

Abstract. In many Semantic Web domains a tremendous number of statements (expressed as triples) can potentially be true but, in a given domain, only a small number of statements is known to be true or can be inferred to be true. It thus makes sense to attempt to estimate the truth values of statements by exploring regularities in the Semantic Web data via machine learning. Our goal is a “push-button ” learning approach that requires a minimum of user intervention. The learned knowledge is materialized off-line (at loading time) such that querying is fast. We define an extension of SPARQL for the integration of the learned probabilistic statements into querying. The proposed approach deals well with typical properties of Semantic Web data. i.e., with the sparsity of the data and with missing data. Statements that can be inferred via logical reasoning can readily be integrated into learning and querying. We study learning algorithms that are suitable for the resulting high-dimensional sparse data matrix. We present experimental results using a friend-of-a-friend data set. 1