On the design of metric relations.

摘要: Metric distances and the more general concept of dissimilarities are widely used tools in instance-based learning methods very especially nearestneighbor classification technique. This paper contributes to design dissimilarity measures increase their utility. The ability understand main properties a hand-crafted measure alter them if necessary greatly widens its applicability. Grounded upon formal definition for measure, together with set fundamental properties, body results is presented concerning equivalence between dissimilarities. Moreover, important concepts transitivity aggregation studied. Results preserving under transformations and/or aggregations presented, emphasis relationship transitivity. Further, issue dealing special values (e.g. missing values) Although focus dissimilarities, particular case metric (i.e. fulfilling triangular inequality) specifically covered, given importance. Several examples use potential utility worked out throughout text