Morik, K., Brockhausen, P. (1996) A Multistrategy Approach to Relational Knowledge Discovery in Databases. Proceedings of the Third International Workshop on Multistrategy Learning (MSL-96), pp. 17-27  Home/Search   Document Details and Download   Summary   Related Articles   Check

....access to all system operators and components. Researchers have started to acknowledge the advantages in integrating inductive inference capabilities in a database system. Among the approaches somewhat similar in philosophy to the one presented here are ones presented by Morik and Brockhausen [14], Mannila [7] and Roddick and Rice [15] In general, VINLEN di#ers from these in terms of its much wider variety and complexity of knowledge generation operators, its knowledge storage mechanism, and its visual interface. Fig. 1 presents a general schema for VINLEN. The top part presents a ....

Morik, K., Brockhausen, P. (1996) A Multistrategy Approach to Relational Knowledge Discovery in Databases. Proceedings of the Third International Workshop on Multistrategy Learning (MSL-96), pp. 17-27

....In the Mining Mart architecture, machine learning algorithms can be used as pre processing operators. For multi relational data sets, it is advantageous to use ILP algorithms, which are able to handle several relations at once. Whereas some ILP algorithms are able to learn on databases directly [Morik and Brockhausen, 1997], others are unable to make use of a relational database. Looking at new developments in ILP like CILGG, which provides a very rich and expressive representation language (a combination of horn logic with description logic, see chapter 4) it is rather unlikely that these kinds of algorithms will ....

....learned by ILP algorithms on excerpts of the data. Both kinds of information, the low level information about physical objects and the high level information about concepts, relations, and features is stored in M 4 and can be read automatically. To reach our second goal and in contrast to RDT DB [Morik and Brockhausen, 1997], where the user has to specify a mapping directly, here, we have chosen an approach where the user specifies the mapping rather indirectly by carrying out pre processing on tables. This gives us a much larger flexibility and at the same time simplifies the mapping itself a lot. Since we have an ....

Morik, Katharina and Brockhausen, Pe- ter, A Multistrategy Approach to Relational Knowledge Discovery in Databases, Machine Learning Journal volume 27, June 3 1997, pages 287312, Kluwer, 1997. 33

....predicates, where the database key is one of the arguments of the predicate. A tool that gives an ILP learner direct access to a relational database, provides different types of mappings from tables to predicates, and constructs SQL queries for hypothesis testing automatically has been developed [17]. However, the open question remains whether there are theoretically well based indicators of the optimal mapping for a particular learning task. This is an essential question, since the number of possible mappings is only bounded by the size of the universal database relation. Therefore, it does ....

Katharina Morik and Peter Brockhausen. A multistrategy approach to relational knowledge discovery in databases. Machine Learning Journal, 27(3):287--312, jun 1997.

....are learned from very large and high dimensional data sets are not interesting, but are already known to the users. For instance, we have learned from a very large data set on cars and their warranty cases that the production date of a car is preceding its sales date, which is what we expected [14]. However, there were some exceptions to the rule and these are interesting. Which customers order cars before they are offered Are the exceptions typing errors The decision can only be made by domain experts. Either we use the outliers for data cleaning or we find interesting instances by ....

Katharina Morik and Peter Brockhausen. A multistrategy approach to relational knowledge discovery in databases. Machine Learning Journal, 27(3):287--312, jun 1997.

....we are neither concerned with the direction of interventions nor the time points when to intervene. Instead, and in the spirit of discovery science, our interest is to find evidence in favour or against the hypothesis mentioned above. For all learning experiments we used the learning system Rdt Db (Morik and Brockhausen, 1997), which is a variation of Rdt (Kietz and Wrobel, 1992) Rdt Db uses for hypothesis generation the core of Rdt, but for hypothesis testing, it translates hypotheses into SQL queries, which are executed and evaluated by an Oracle database. Rdt Db uses a declarative, syntactic bias for learning, ....

Morik, K. and Brockhausen, P. (1997). A Multistrategy Approach to Relational Knowledge Discovery in Databases. Machine Learning Journal, 27(3):287--312.

....engine of MOBAL derives expected effects and compares them with actual effects. These are deductive inferences. However, the explanation of conflicts between prediction and actual outcome requires to investigate many hypotheses. For this task, we used the inductive logic programming tool RDT DB (Morik and Brockhausen, 1997). Work on action effect rules and validation is described in section 4. Data Set. The data set was collected at the 16 bed intensive care unit (ICU) of the Chirurgische Kliniken der Stadtischen Kliniken Dortmund . It contains the data of 147 patients between January 1997 and October 1998. ....

Morik, K. and Brockhausen, P. (1997). A Multistrategy Approach to Relational Knowledge Discovery in Databases. Machine Learning Journal, 27(3):287--312.

....of or the whole key in any of the above mappings. Having a table with three attributes A; B; and C, A is the key, the user might be only interested in the different combinations of 2 The learning algorithm Num Int discovers intervals of numerical values, based on a gap approach, for details cf. [18]. B and C. Then he can specify a predicate combinations(B; C) Using this predicate in a hypothesis, the Sql generator takes care of three different things. First, a group by B, C statement will be inserted into the query. Second, a projection onto B and C will be build, and third, only ....

....Fdd (Functional Dependency Detection) 3] finds a generality order of attributes by detecting functional dependencies in the database. Num Int finds a hierarchy of intervals in numerical (linear) attributes without reference to a classification. This approach is discussed in detail in [18]. In the following we concentrate on another alternative, namely the use of background knowledge. We will demonstrate both, that we can use it in an appropriate way, and that we need it to learn more interesting rules. 5 Note that neither the rule nor the percentages involved are the true ones ....

Katharina Morik and Peter Brockhausen. A multistrategy approach to relational knowledge discovery in databases. In Ryszard S. Michalski and Janusz Wnek, editors, Proceedings of the Third International Workshop on Multistrategy Learning (MSL-96), Palo Alto, May 1996. AAAI Press.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC