Semantic Web knowledge bases, in particular large cross-domain data, are often noisy, incorrect, and incomplete with respect to type information. This incompleteness can be reduced, as previous work shows, with automatic type prediction methods. Most knowledge bases contain an ontology defining a type hierarchy, and, in general, entities are allowed to have multiple types (classes of an instance assigned with the rdf:type relation). In this paper, we exploit these characteristics and formulate the type prediction problem as hierarchical multi classification, where the labels are types. We evaluate different sets of features, including entity embeddings, which can be extracted from the knowledge graph exclusively. We propose SLCN, a modification of the local classifier per node approach, which performs feature selection, instance sampling, and class balancing for each local classifier with the objective of improving scalability. Furthermore, we explore different variants of creating features for the classifier, including both graph and latent features. We compare the performance of our proposed method with the state-of-the-art type prediction approach and popular hierarchical multilabel classifiers, and report on experiments with large-scale cross-domain RDF datasets.
Type Prediction in Noisy RDF Knowledge Bases Using Hierarchical Multilabel Classification with Graph and Latent Features
