Petr Berka

Discretization and Grouping: Preprocessing Steps for Data Mining

Unlike on-line discretization performed by a number of machine learning (ML) algorithms for building decision trees or decision rules, we propose off-line algorithms for discretizing numerical attributes and grouping values of nominal attributes. The number of resulting intervals obtained by discretization depends only on the data; the number of groups corresponds to the number of classes. Since both discretization and grouping is done with respect to the goal classes, the algorithms are suitable only for classification/prediction tasks.

Data Mining and Medical Knowledge Management: Cases and Applications

The healthcare industry produces a constant flow of data, creating a need for deep analysis of databases through data mining tools and techniques resulting in expanded medical research, diagnosis, and treatment. Data Mining and Medical Knowledge Management: Cases and Applications presents case studies on applications of various modern data mining methods in several important areas of medicine, covering classical data mining methods, elaborated approaches related to mining in electroencephalogram and electrocardiogram data, and methods related to mining in genetic data. A premier resource for those involved in data mining and medical knowledge management, this book tackles ethical issues related to cost-sensitive learning in medicine and produces theoretical contributions concerning general problems of data, information, knowledge, and ontologies.

Empirical Comparison of Various Discretization Procedures

The genuine symbolic machine learning (ML) algorithms are capable of processing symbolic, categorial data only. However, real-world problems, e.g. in medicine or finance, involve both symbolic and ...

Discovering Company Descriptions on the Web by Multiway Analysis

We investigate the possibility of web information discovery and extraction by means of a modular architecture analysing separately the multiple forms of information presentation, such as free text, structured text, URLs and hyperlinks, by independent knowledge-based modules. First experiments in discovering a relatively easy target, general company descriptions, suggests that web information can be efficiently retrieved in this way. Thanks to the separation of data types, individual knowledge bases can be much simpler than those used in information extraction over unified representations.

AtherEx: an expert system for atherosclerosis risk assessment

A number of calculators that compute the risk of atherosclerosis has been developed and made available on the Internet. They all are based on computing weighted sum of risk factors. We propose instead to use more flexible expert systems to estimate the risk. The goal of the AtherEx expert system is to classify patients according to their atherosclerosis risk into four groups. This application is based on the Nest rule–based expert system shell. Knowledge for the AtherEx was obtained (using the machine learning algorithm KEX) from the data concerning a longitudial study of atherosclerosis risk factors and further refined by domain expert. AtherEx is available for consultations on web.

Recognizing Reliability of Discovered Knowledge

When using discovered knowledge for decision making (e.g. classification in the case of machine learning), the question of reliability becomes very important. Unlike global view on the algorithms (evaluation of overall accuracy on some testing data) or unlike multistrategy learning (voting of more classifiers), we propose “local” evaluation for each example using one classifier. The basic idea is to learn to classify the correct decisions made by the classifier. This is done by creating new class attribute “match” and by running the learning algorithm on the same input attributes. We call this (second) step verification. Some first preliminary experimental results of this method used with C4.5 and CN4 are reported. These results show that: (1) if the classification accuracy is very high, it makes no sence to perform the verification step (since the verification step will create only the majority rule), (2) in multiple-class and/or noisy domains the verification accuracy can be significantly higher then the classification accuracy.

Predicting page occurrence in a click-stream data: statistical and rule-based approach

We present an analysis of the click-stream data with the aim to predict the next page that will be visited by an user based on a history of visited pages. We present one statistical method (based on Markov models) and two rule induction methods (first based on well known set covering approach, the other base on our compositional algorithm KEX). We compare the achieved results and discuss interesting patterns that appear in the data.

Atherosclerosis Risk Assessment using Rule-Based Approach

A number of calculators that compute the risk of atherosclerosis has been developed and made available on the Internet. They all are based on computing weighted sum of risk factors. We propose instead to use more flexible rule-based approach to estimate this risk. The used rules were created using machine learning methods and further refined by domain expert. Using our rule-based expert system NEST, we built a consultation module AtherEx, that helps (via Internet) a non-expert user to evaluate his atherosclerosis risk.

Rule Induction for Click-Stream Analysis: Set Covering and Compositional Approach

We present a set covering algorithm and a compositional algorithm to describe sequences of www pages visits in click-stream data. The set covering algorithm utilizes the approach of rule specialization like the well known CN2 algorithm, the compositional algorithm is based on our original KEX algorithm, however both algorithms deal with sequences of events (visited pages) instead of sets of attributevalue pairs. The learned rules can be used to predict next page to be viewed by a user or to describe the most typical paths of www pages visitors and the dependencies among the www pages. We have successfully used both algorithms on real data from an internet shop and we mined useful information from the data.

Adaptive features of machine learning methods

This paper gives a survey of (symbolic) machine learning methods, that exhibit significant features of adaptivity. The paper discusses incremental learning, learning in dynamically changing domains, knowledge integration, theory revision, case based reasoning and inductive logic programming.