From OpenCog


Finally, a different method for creating new ConceptNodes in OCP is using clustering algorithms. There are many different clustering algorithms in the statistics and data mining literature, and no doubt many of them could have value inside OCP. We have experimented with several different clustering algorithms in the OCP context, and have selected one, which we call Omniclust, based on its generally robust performance on high-volume, noisy data.

In the discussion on OpenCogPrime:EvolutionaryConceptCreation, we mentioned the use of a clustering algorithm to cluster links. The same algorithm we describe here for clustering ConceptNodes directly and creating new ConceptNodes representing these clusters, can also be used for clustering links in the context of node mutation and crossover.

The application of Omniclust or any other clustering algorithm for ConceptNode creation in OCP is simple. The clustering algorithm is run periodically, and the most significant clusters that it finds are embodied as ConceptNodes, with InheritanceLinks to their members. If these significant clusters have subclusters also identified by Omniclust, then these subclusters are also made into ConceptNodes, etc., with InheritanceLinks between clusters and subclusters.

Clustering technology is famously unreliable, but this unreliability may be mitigated somewhat by using clusters as initial guesses at concepts, and using other methods to refine the clusters into more useful concepts. For instance, a cluster may be interpreted as a disjunctive predicate, and a search may be made to determine sub-disjunctions about which interesting OpenCogPrime:PLN conclusions may be drawn.