Karol Grudzinski

Learn More
Neural networks and many other systems used for classification and approximation are not able to handle symbolic attributes directly. A method of replacing the symbolic values of attributes by numerical values is described here. The probabilistic Value Difference Metric (VDM) and similar functions used in the Nearest-Neighbor algorithms in most cases work(More)
SBL-PM is a simple algorithm for selection of reference instances, a first step towards building a partial memory learner. A batch and on-line version of the algorithm is presented, allowing to find a compromise between the number of reference cases retained and the accuracy of the system. Preliminary experiments on real and artificial datasets illustrate
Analysis of medical data requires not only classification of patterns but also some data understanding. Several systems for extraction of logical rules from data have been applied to analysis of the melanoma skin cancer data. These systems include neural, decision tree and inductive algorithms for rule extraction and minimal-distance methods used for(More)
Framework for Similarity-Based Methods (SBMs) allows to create many algorithms that differ in important aspects. Although no single learning algorithm may outperform other algorithms on all data an almost optimal algorithm may be found within the SBM framework. To avoid tedious experimentation a meta-learning search procedure in the space of all possible(More)
Several methods for feature selection and weighting have been implemented and tested within the similarity-based framework of classification methods. Features are excluded and ranked according to their contribution to the classification accuracy in the crossvalidation tests. Weighting factors used to compute distances are optimized using global minimization(More)
As a step towards neural realization of various similarity based algorithms k-NN method has been extended to weighted nearest neighbor scheme. Experiments show that for some datasets significant improvements are obtained. As an alternative to the minimization procedures a best–first search weighted nearest neighbor scheme has been implemented. A feature(More)
Attempts to extract logical rules from data often lead to large sets of classification rules that need to be pruned. Training two classifiers, the C4.5 decision tree and the Non-Nested Generalized Exemplars (NNGE) covering algorithm, on datasets that have been reduced earlier with the EkP instance compressor leads to statistically significantly lower number(More)