The broad goal of knowledge discovery in databases (KDD) is, fittingly, to construct knowledge from large spatial database systems (SDBS). This goal is achieved via spatial data mining methods (algorithms) which are used to automate KDD tasks (e.g. detection of classes, dependencies, anomalies). Without a fuller understanding of the field at present, it is hard to judge how comprehensive an approach is outlined in Ester et al’s (1997) paper.
The authors underline the distinguishing characteristics of spatial databases; namely, the assumption that an object’s attributes may be influenced by the attributes of its neighbours (Tobler). These assumptions motivate the development of techniques and algorithms which automate the identification and extraction of spatial relationships. For instance, a simple classification task can be executed by algorithms that group objects based on the value of their attributes. The authors present a spatial extension of this approach, by incorporating not only an object’s attributes, but also those of its neighbours, allowing for greater insight into spatially classified sets of objects within a SDBS.
Contrasting with last week’s topic, the approach to knowledge extraction here emphasises automation. The goal is to construct basic rules that can efficiently manipulate and evaluate large datasets to detect meaningful, previously unknown information. Certainly, these techniques have been invaluable for pre-processing, transforming, mining and analysing large databases. In light of recent advances, it would be interesting to revisit these techniques to assess whether new spatial data mining methods are more effective for guessing or learning patterns that may be interpreted as meaningful, and to consider the theoretical limits of these approaches (if they exist).
-slumley