Sinh Hoa Nguyen

Rough set algorithms in classification problem

We we present some algorithms, based on rough set theory, that can be used for the problem of new cases classification. Most of the algorithms were implemented and included in Rosetta system [43]. We present several methods for computation of decision rules based on reducts. We discuss the problem of real value attribute discretization for increasing the performance of algorithms and quality of decision rules. Finally we deal with a problem of resolving conflicts between decision rules classifying a new case to different categories (classes). Keywords: knowledge discovery, rough sets, classification algorithms, reducts, decision rules, real value attribute discretization

Layered learning for concept synthesis

We present a hierarchical scheme for synthesis of concept approximations based on given data and domain knowledge. We also propose a solution, founded on rough set theory, to the problem of constructing the approximation of higher level concepts by composing the approximation of lower level concepts. We examine the effectiveness of the layered learning approach by comparing it with the standard learning approach. Experiments are carried out on artificial data sets generated by a road traffic simulator.

Regularity analysis and its applications in data mining

Abstract: Knowledge discovery is concerned with extraction of useful information from databases ([21]). One of the basic tasks of knowledge discovery and data mining is to synthesize the description of some subsets (concepts) of entities contained in databases. The patterns and/or rules extracted from data are used as basic tools for concept description. In this Chapter we propose a certain framework for approximating concepts. Our approach emphasizes extracting regularities from data. In this Chapter the following problems are investigated: (1) issues concerning the languages used to represent patterns; (2) computational complexity of problems in approximating concepts; (3) methods of identifying, optimal patterns. Data regularity is a useful tool not only for concept description. It is also indispensable for various applications like classification or decomposition. In this Chapter we present also the applications of data regularity to three basic problems of data mining: classification, data description and data decomposition.

Discovery of Data Patterns with Applications to Decomposition and Classification Problems

Data mining community is searching for efficient methods of extracting patterns from data [20],[22],[39],[46],[45]. We study problems of extracting several kinds of patterns from data. The simplest ones are called templates. We consider also more sophisticated relational patterns extracted automatically from data.

Pattern Extraction from Data

Searching for patterns is one of the main goals in data mining. Patterns have important applications in many KDD domains like rule extraction or classification. In this paper we present some methods of rule extraction by generalizing the existing approaches for the pattern problem. These methods, called partition of attribute values or grouping of attribute values, can be applied to decision tables with symbolic value attributes. If data tables contain symbolic and numeric attributes, some of the proposed methods can be used jointly with discretization methods. Moreover, these methods are applicable for incomplete data. The optimization problems for grouping of attribute values are either NP-complete or NP-hard. Hence we propose some heuristics returning approximate solutions for such problems.

On Finding Optimal Discretizations for Two Attributes

We show that finding optimal discretization of instances of decision tables with two attributes with real values and binary decisions is computationally hard. This is done by abstracting the problem in such a way that it regards partitioning points in the plane into regions, subject to certain minimality restrictions, and proving them to be NP-hard. We also propose a new method to find optimal discretizations.

Rough Set Methods in Approximation of Hierarchical Concepts

Many learning methods ignore domain knowledge in synthesis of concept approximation. We propose to use hierarchical schemes for learning approximations of complex concepts from experimental data using inference diagrams based on domain knowledge. Our solution is based on the rough set and rough mereological approaches. The effectiveness of the proposed approach is performed and evaluated on artificial data sets generated by a traffic road simulator.

Rough sets and association rule generation

ASSOCIATION RULE (see [1]) extraction methods have been developed as the main methods for mining of real life data, in particular in Basket Data Analysis. In this paper we present a novel approach to generation of association rules, based on Rough Set and Boolean reasoning methods. Some results presented in this paper has been mentioned in [13, 17]. We will explain them precisely (with full proofs of theorems) in this paper. We show the relationship between the problems of association rule extraction for transaction data and relative reducts (or α-reducts generation) for a decision table. Moreover, the present approach can be used to extract association rules in general form. The experimental results show that the presented methods are quite efficient. Large number of association rules with given support and confidence can be extracted in a short time.

Rough set approach to sunspot classification problem

This paper presents an application of hierarchical learning method based rough set theory to the problem of sunspot classification from satellite images. The Modified Zurich classification scheme [3] is defined by a set of rules containing many complicated and unprecise concepts, which cannot be determined directly from solar images. The idea is to represent the domain knowledge by an ontology of concepts – a treelike structure that describes the relationship between the target concepts, intermediate concepts and attributes. We show that such ontology can be constructed by a decision tree algorithm and demonstrate the proposed method on the data set containing sunspot extracted from satellite images of solar disk.

Searching for Relational Patterns in Data

We consider several basic classes of tolerance relations among objects. These (global) relations are defined from some predefined similarity measures on values of attributes. A tolerance relation in a given class of tolerance relations is optimal with respect to a given decision table A if it contains only pairs of objects with the same decision and the number of such pairs contained in the relation is maximal among all relations from the class. We present a method for (sub-)optimal tolerance relation learning from data (decision table). The presented method is based on rough set approach. We show that for some basic families of tolerance relations this problem can be transformed to a relative geometrical problem in a real affine space. Hence geometrical computations are becoming useful tools for solving the problem of global tolerance relation construction. The complexity of considered problems can be evaluated by the complexity of the corresponding geometrical problems. We propose some efficient heuristics searching for an approximation of optimal tolerance relations in considered families of tolerance relations. The global tolerance relations can be treated as patterns in the cartesian product of the object set. We show how to apply the relational patterns (global tolerance relations) in clustering and classification of objects.