Search

Genetic algorithms have been utilized in many complex optimization and simulation tasks because of their powerful search method. In this research we studied whether the classification performance of the attribute weighted methods based on the nearest neighbour search can be improved when using the genetic algorithm in the evolution of attribute weighting. The attribute weights in the starting population were based on the weights set by the application area experts and machine learning methods instead of random weight setting. The genetic algorithm improved the total classification accuracy and the median true positive rate of the attribute weighted k-nearest neighbour method using neighbour’s class-based attribute weighting. With other methods, the changes after genetic algorithm were moderate.

Data mining is the process of getting useful information by analyzing different kind of data. Predictive data mining is used to predict some property of incoming data for example how to classify it. Among many methods that are used for predictive data mining the K-nearest neighbor classification is one of the simplest and easy to use technique. Due to its simplicity small variations are possible with it for the purpose of improving its predictive accuracy. The aim of this thesis was to study attribute weighting techniques and to implement and test some weighting variants in K-nearest neighbor classification. The HEOM distance metric and three values of K (1, 4 and 5) were used in K-nearest neighbor classification. Twelve datasets were selected from the UCI Machine Learning Repository for the analysis. Chi-square attribute weighting was done in order to implement the two weighting variants. One variation was the simple attribute weighting and the other was the class-wise attribute weighting. The evaluation was done by using the leave-one-out technique. The conclusion that can be drawn from the results is that the structure of the dataset (the number and the distribution of the classes) and the value of K (the number of neighbors) have effect on the unweighted and attribute weighted K-nearest neighbor classification. For some datasets weighting is very useful especially for smaller classes, but for some datasets it does not give improvements in the result.

How do actors involved in the acquisition of capital goods assess the importance of their attributes? What is the role of expertise? Numerous instruments exist for measuring the importance attached to attributes, but little is known about the importance assessment process that precedes these importance judgments. Expectations concerning the behaviour of actors facing non-routine importance assessment problems are tested, yielding some interesting results. Firstly, the behaviour of these actors is consistent with a newly developed phase model. Even with a non-routine problem, structuring the assessment problem takes less effort than the actual weighting. Surprisingly, weighting attributes in isolation gets much more emphasis than weighting them against each other, despite the latter being the essence of importance judgments. Despite the subjects being laymen, they showed high confidence in their work. Finally, predictions concerning the behaviour of experts are made, based on Van der Heijden's dimensions of expertise.

We propose a binary classifier based on the single hidden layer feedforward neural network (SLFN) using radial basis functions (RBFs) and sigmoid functions in the hidden layer. We use a modified attribute-class correlation measure to determine the weights of attributes in the networks. Moreover, we propose new weights called as influence weights to utilize in the weights connecting the input layer and the hidden layer nodes (hidden weights) of the network with sigmoid hidden nodes. These weights are calculated as the sum of conditional probabilities of attribute values given class labels. Our learning procedure of the networks is based on the extreme learning machines; in which the parameters of the hidden nodes are first calculated and then the weights connecting the hidden nodes and output nodes (output weights) are found. The results of the networks with the proposed weights on some benchmark data sets show improvements over those of the conventional networks.