Premium
Probabilistic tests and stopping rules associated with hierarchical classification techniques
Author(s) -
SANDLAND R. L.,
YOUNG P. C.
Publication year - 1979
Publication title -
australian journal of ecology
Language(s) - English
Resource type - Journals
eISSN - 1442-9993
pISSN - 0307-692X
DOI - 10.1111/j.1442-9993.1979.tb01567.x
Subject(s) - multinomial distribution , classification rule , probabilistic logic , early stopping , computer science , statistics , sampling (signal processing) , mathematics , artificial intelligence , machine learning , data mining , filter (signal processing) , artificial neural network , computer vision
A test of random noise and an objective stopping rule are derived for use in association with hierarchical classification techniques. The apparatus required is replicated sampling and combinatorial analysis. The test and rule are based on the conditional probabilities of the number of sites all of whose replicates belong to the same group at each stage of the classification. These devices were tested on presence/absence data simulated from dependent multinomial trials. Where the number of groups became large the classification techniques proved somewhat unreliable because the groups were too similar. Usually clearer distinctions between groups will occur in nature. However, the test for random noise proved entirely successful and the stopping rule was quite successful, although, as one might expect, it was less successful when the classification itself was less successful.