marco@work - A survey of interestingness measures for knowledge discovery

Ik zou willen dat ik A survey of interestingness measures for knowledge discovery door Ken McGarry uit 2005 een paar maanden eerder had ontdekt... Dit paper evalueert de huidige stand van zaken in het wetenschappelijke onderzoek naar het meten van de mate van interessantheid van associaties tussen variabelen. Deze informatie is voor mij relevant aangezien ik op zoek ben naar afhankelijkheden tussen de syntactische variabelen in de SAND...

Gekopieerd uit de Conclusies...
"This paper has reviewed and evaluated the current research literature on the various techniques for determining the interestingness of patterns discovered by the data mining process. The characteristics of several goal- and data-driven measures have been discussed. The main disadvantage of the subjective or user-driven approach is that it constrains the discovery process to seek only what the user can anticipate or hypothesize, i.e. it cannot discover unexpected or unforeseen patterns because it is entirely goal driven. Alternatively, objective measures or data-driven measures tend to concentrate on finding patterns through statistical strength or correlations. Such patterns may not be interesting to the user as a strong pattern may imply knowledge that is already well known to the organization. We have discussed the issues regarding the degree of user involvement within the knowledge discovery process. The planned level of autonomy will have a critical impact on how the interestingness measures are applied and therefore the overall efficiency and usefulness of the discovery system. A major research question is how to combine both objective and subjective measures into a unified measure. The unification is necessary as developing subjective measures is quite expensive because of the costs associated with the knowledge acquisition process. It is likely that ontologies and other semantic technologies will play an increasingly important role in bridging this gap, by providing the level of detail required to develop robust belief systems. Without a sufficiently rich belief system it is difficult for subjective measures to accurately model the user domain. Recent work in paradox detection has seen the development of interestingness measures that are part way between objective and subjective measures, and perhaps they may provide the foundations for general purpose automated discovery systems."