The centroid or consensus of a set of objects with qualitative attributes

Semantically-grounded construction of centroids for datasets with textual attributes

“…In ontology/taxonomy-based approaches, on the contrary, the search space can be potentially extended to all concepts modelled in the ontology (e.g., taxonomical subsumers of terms contained in the cluster) and, hence, the centroid can be synthetically constructed from a finer grained set of candidates. This advantage, however, is either slightly considered in related works [1], restricting the centroid to the LCS, or even not exploited at all [22,23], in cases in which the background taxonomy only incorporates terms found in the input dataset. To expand the search space and obtain fine-grained centroids, we propose relying on detailed ontologies like WordNet [19].…”

“…In fact, an investigation of the structure of existing ontologies via the Swoogle ontology search engine [12] has shown that available ontologies very occasionally model non-taxonomic knowledge. By analysing taxonomic relationships we retrieve new concepts (i.e., taxonomical ancestors) that become centroid candidates for the values in V. Following a similar premise as in [1,22], we assume that taxonomical subsumers of a term (including itself) are valid representatives of the term. The set of candidates is given in the minimum subsumer hierarchy (H O (V)) that goes from the concepts corresponding to the values in V to the Least Common Subsummer (LCS) of all values.…”

“…In [22,23] authors propose a generic way to calculate the centroid, called consensus, of a set of textual values. Their proposal exploits the knowledge modelled in ad hoc hierarchies that taxonomically link input values to measure their confusion (i.e., a measure of semantic alikeness).…”

“…On the contrary to most related works (discussed in Section 2), which bound the centroid to values already contained in the input dataset [16,22,46], we have put special efforts in maximising the flexibility of the centroid construction. In our approach, we exploit, as much as possible, the knowledge provided by the background ontology, resulting in finer grained and more accurate centroids.…”

“…Hence, when constructing the centroid of a dataset with textual attributes, the similarities between their meanings (evaluated at a conceptual level) should be taken into consideration (e.g., for hobbies attribute, ''trekking'' value is more similar to ''jogging'' than to ''classical dance''). Related works constructing centroids for textual attribute values typically omit [16,46] or slightly consider [1,22] data semantics during their analysis, hampering the quality of the results [6,33].…”