Learning to detect abnormal semantic web data

Abstract

Numerous problems could happen in the generation process for Semantic Web data that is usually gathered from heterogeneous sources by using a variety of tools [3]. Recently some works [1, 2, 3, 4] began to focus on the quality of Semantic Web data. However since the Semantic Web represents many points of view, there is no objective measure of correctness for all Semantic Web data. Therefore, we consider using an abnormality heuristic that could indicate a data quality problem at the triple level. We recognize that not all abnormal data is incorrect (in fact, in some scenarios the abnormal data may be the most interesting data) and thus leave it up to the application to determine how to use the heuristic. The essential idea of this work is based on the fact that a statement can get supporting evidence if it can be entailed from other data. Consider the statement A advises B: in some situations where this is true, there are also statements such as A is the principal investigator of project C, B works in C. This rule is clearly not certain. Yet, when combined with other forms of evidence, it can provide support for the advises relation. To detect incorrect data, ideally we can directly learn characteristics of them. But incorrect data have too many forms. So we check if the data lacks sufficient normal patterns compared to the majority of the data. Still using the advises relation example above, we change the first statement into B advises A (assuming advises is not subPropertyOf advises). Then our predictability on this statement would be low, because the context is inconsistent with a probabilistic rule existing in many other contexts. Although this probabilistic rule does not always hold, various rules in context can collaboratively give certain support. Note that there are many possible arbitrary relations that can be used to describe any two objects on the Semantic Web, but the