Grete Lind

An Effective Inductive Learning Algorithm for Extracting Rules

In this paper we present a new inductive learning algorithm named MONSAMAX for extracting rules. It has some advantages compared to several machine learning algorithms: it uses several new pruning techniques which guarantee great effectiveness of the algorithm; it extracts overlapping rules; as a result it finds determinative set of rules that we can use for post-analysis of extracted rules. Compared to a former algorithm MONSIL it is much less labor-consuming.

Machine Learning Task as a Diclique Extracting Task

As we know there exist several approaches and algorithms for data mining and machine learning task solution, for example, decision tree learning, artificial neural networks, Bayesian learning, instance-based learning, genetic algorithms, etc. They are effective and well-known and their base algorithms and main ideology are published. In this paper we present a new approach for machine learning (ML) task solution, an inductive learning algorithm based on diclique extracting task. We show how to transform ML as inductive leaning task into the graph theoretical diclique extracting task, present an example and discuss about the problems related with that approach and effectiveness of the algorithm.

Algorithm for Finding Zero Factor Free Rules

Class detection rules are mostly used for classifying new objects. Another possible usage is to describe a set of objects (a class) by the rules. Determinacy Analysis (DA) is a knowledge mining method with such purpose. Sets of rules are used to answer the questions “Who are they (objects of the class)?”, “How can we describe them?”. Rules found by different DA methods tend to contain some redundant information called zero factors. In this paper we show how zero factors are related to closed sets and minimal generators. We propose a new algorithm that extracts zero-factor-free rules and zero factors themselves, based on finding generators. Knowing zero factors gives to the analyst important additional knowledge for understanding the essence of the described set of objects (a class).

Some developments of determinacy analysis

This paper deals with development of Determincy Analysis (DA), a method for data mining. There are two approaches to DA that give a different result consisting of non-overlapping (exclusive each other) rules. The first method finds exactly one system consisting of rules that have equal length. The second, step by step approach enables to find very many different rule systems where the rules have different length. In the first case the rules contain a lot of redundant attributes, in the second case there are too many different (formally complete) systems of rules what makes the selection hard. A better result can be obtained by finding overlapping rules. This paper presents DA approach and technique that enables to find overlapping rules with different length and algorithm realizing it. Such approach for DA has not been created earlier.