A multi-mechanism rule-extraction pipeline for use on unannotated datasets

We outline a hybrid methodology (incorporating both supervised and unsupervised-learning components) for rule-based knowledge discovery from unannotated data i.e. when the classification information is unknown. The motivation for our work stems from the individual effectiveness of various data mining mechanisms i.e.: (1) class identification via unsupervised datavector cluster formation, (2) datavector simplification and feature selection via attribute discretisation, and (3) symbolic rule extraction via the association of symbolic rules with the structural parameters of a trained neural network (NN). The basic operational concept involves the pipelined application of various unsupervised and supervised mechanisms i.e.: (1) k-means, (2) Chi-2, (3) local cluster (LC) network training, and (4) rule extraction from a trained LC network. The methodology will be tested and analysed using several well-known datasets