Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
530721 | Pattern Recognition | 2012 | 11 Pages |
This paper presents an on-line multi-stage sorting algorithm capable of adapting to different populations. The sorting algorithm selects on-line the most appropriate classifier and feature subsets for the incoming population. The sorting algorithm includes two levels, a low level for population detection and a high level for classifier selection which incorporates feature selection. Population detection is achieved by an on-line unsupervised clustering algorithm that analyzes product variability. The classifier selection uses n fuzzy kNN classifiers, each trained with different feature combinations that function as input to a fuzzy rule-based decision system. Re-training of the n fuzzy kNN classifiers occurs when the rule based system cannot assign an existing classifier with high confidence level. Classification results for synthetic and real world databases are presented.
► On-line multi-stage sorting algorithm capable of adapting to different populations. ► One level for population detection and another level for classifier selection. ► Population detection is achieved by an on-line unsupervised clustering algorithm. ► Classifier selection using fuzzy kNN classifiers, trained with different features. ► The fuzzy kNN classifiers are retrained when an existing classifier is not assigned.