PATTERN-RECOGNITION FOR SAMPLE CLASSIFICATION USING ELEMENTAL COMPOSITION - APPLICATION FOR INDUCTIVELY-COUPLED PLASMA-ATOMIC EMISSION-SPECTROMETRY

Three pattern recognition techniques were investigated as tools for au tomatic recognition of samples: k-Nearest Neighbors, Bayesian Classifi cation and the C4.5 inductive learning algorithm. Their abilities to c lassify 20 geological reference materials were compared. Each training and test example used 13 elemental concentrations. The data set was c omposed of 2582 examples obtained from CANMET in the form of results o f analyses performed on these reference materials by different laborat ories. It was found that all three pattern recognition techniques Perf ormed extremely well with a large data set of real samples. Bayesian C lassification and k-Nearest Neighbors worked very well with small data sets.