CB1D - Clustering Based 1D Learner

The CB1D learner was developed as the response to the task given at Automated Learning lectures at the University of Amsterdam, Netherlands in the school year 2005/2006. The lecture was given by Maarten van Someren.

Task description

Purpose of this assignment is to invent a method, develop a system and justify this, for predicting a class from a single numerical variable. Training data are a set of cases, where a case consists of a single value and a class label. There are only two classes (called + and -).

Data set

The used data set is available from http://www.ics.uci.edu/~mlearn/databases/credit-screening/. It describes, if someone gets a loan or not. The data come from real financial institution, only the labels where changed in order not to reveal sensitive information.

There are several attributes available. It is allowed to select any single continuous variable to make the +/- classification.

Motivation

This seemingly simple objective allows for simple and straightforward methods evaluation. However, the noisy data make the problem "real-world-like".

Algorithm - Extension of RBF neural networks

Because the provided data have highly local patterns, the choice was made to use Radial Basis Function neural networks. However, this kind of network is widely used for function interpolation, not for classification. Therefor the general RBF learning algorithm was modified to provide better results.

Final paper

The final report is available for download: jiri.isa_clustering.based.1D.learner.pdf. This paper introduces reader to the task, describes the general RBF learning algorithm and the CB1D improvement. Finally it compares the same implementation of classical RBF with CB1D to show, that the improvement indeed gains some improvement.