Learning of Inexact Rules by the Fish-net Algorithm from Low Quality Data

We present an algorithm, the FISH-NET algorithm, for deriving classiica-tion/forecasting rules from large data bases of low quality data. The attributes are assumed to be continuous, numeric variables. The algorithm works on the eld of the attributes, rather than on individual point values and is linear in both the number of attributes and the number of instances. The algorithm has been tested on 3 large meteorological data bases. Experimental results show that the algorithm outperforms C4.5, feed forward neural networks, k-nearest neighbour classiiers and human experts on these data sets.