The PKDD Discovery Challenges on Thrombosis Data

The aim of the Discovery Challenge workshops held during PKDD conferences is to encourage a collaborative research effort when analyzing real world data. For PKDD’99 and PKDD2000 two data sets were available; from the financial and from the medical domain, for PKDD2001 only the medical data are used. There are two basic types of contibutions to the Challenge; in “method oriented” papers the authors describe their own approach and use the data mainly for demonstration, in “problem oreinted” papers the authors tried to solve a problem that can be interesting for the end user. 1 Discovery Challenges at the PKDD Conferences The aim of the Discovery Challenge workshops held during PKDD conferences is to encourage a collaborative research effort when analyzing real world data. The idea came from Jan Zytkow, who suggested to organize such an event during PKDD' 99 in Prague. In contrast to competitive nature of KDD Cups held within KDD Conferences, the Discovery Challenge emphasises the aspect of cooperation. Two data sets were available for the PKDD’99 and PKDD2000 Discovery Challenges. In the financial domain, the dataset describes clients of a bank, their accounts, transactions, permanent orders, granted loans and issued credit cards. In the medical domain, the dataset describes patients with collagen diseases. The PKDD2001 Challenge is organized only around the medical data. Each participant could use any KDD techniques and discover as much knowledge as possible. Ideally each contribution includes the proposed business objectives (goals that may be of interest to database users), a brief summary of datamining effort, presentation of the discovered knowledge, and an explanation for database users how they can apply the discovered knowledge. 2 Medical Domain the Thrombosis Data The Thrombosis Data for the PKDD1999 Discovery Challenge were organized into three tables, TSUM_A, TSUM_B, TSUM_C. The tables can be connected by the ID number unique for each patient. Table TSUM_A gives basic information about patients (input by doctors). This dataset includes all patients (about 1000 records). Table TSUM_B gives special laboratory examinations (input by doctors) (measured by the Laboratory on Collagen Diseases). This dataset does not include all the patients, but includes the patients with these special tests. The data in table TSUM_C are data about laboratory examinations stored in Hospital Information Systems (Stored from 1980 to March 1999); all the data include ordinary laboratory examinations and have temporal stamps. The tests are not necessarily connected to thrombosis. For the PKDD2000 Discovery Challenge, the data was restructured into 7 tables to eliminate problems with multi-valued attributes in the original tables (for details see the data description by Zytkow and Gupta in this volume). The same data tables are used also for the challenge this year. 3 The PKDD experience Altogether ten papers on the thrombosis data analysis have been presented at the PKDD’99, PKDD2000 and PKDD2001 conferences. Most of the contributions deal with the classification of thrombosis, but there have also been papers dealing with temporal aspects of the data. We can distinguish two basic types of the contributions. The “method/algorithm oriented” papers focus on describing a new approach or system and use the data more or less for demonstration of the features of the method. The “problem oriented” papers try to formulate (and solve) a problem which can be interesting for end users or domain experts. Tables 1-3 summarize all the papers in terms of solved problem (task), described KDD steps, used mining algorithms and used system. All the papers are available from web at http://lisp.vse.cz/challenge. Table 1. PKDD’99 results 1st. author KDD task KDD steps DM method tool Beilken correlations between lab. Test + thrombosis vizualization Display correlations InfoZoom (own) Levin predict yes/no thrombosis description association rules, ranking objects WizWhy (own) Taylor predict thrombosis, diagnoses preprocessing, classification classification and regression trees Table 2. PKDD 2000 results 1. Author KDD task KDD steps DM method tool Meidan predict yes/no thrombosis description association rules, ranking objects WizWhy (own) Tawfik causal and temporal patterns preprocessing, description, (classification) statistical techniques (Bayesian networks) Tetrad Table 3. PKDD 2001 results 1. Author KDD task KDD steps DM method tool Boulicaut Classify collagen disease preprocessing, description association rules, classification rules ac-miner-12 (own) Coursac classify thrombosis preprocessing, classification decision trees and rules C5.0 Jensen classify thormbosis CRISP-DM neural networks, decision rules, sequnece analysis, association rules Clementine Werner classify severity of disease classification genetic programming LilGP (own) Zytkow classify severity of disease description, classification, interpretation SQL, contingency tables 4 Thrombosis data at another challenges Beside the PKDD conferences, another challenges used the Thrombosis data as well. In September 1999, Shusaku Tsumoto organized a special session in the 38 SIG-FAI and the 45 SIG/KBS of Japanese Society for Artificial Intelligence.