Van Laer, W., De Raedt, L., D zeroski, S.: On Multi-Class Problems and Discretization in Inductive Logic Programming, in: Proceedings of the 10th International Symposium on Methodologies for Intelligent Systems (ISMIS97) (Z. Ra s, A. Skowron, Eds.), vol. 1325 of LNAI, Springer-Verlag, 1997, 277--286.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....strategy has to be adopted for resolving con icts among the rules (e.g. 4] Furthermore, it may very well happen that none of the rules is applicable when trying to classify new examples. A common strategy for handling such examples is to classify them as belonging to the majority class (e.g. [7]) In this paper we present a new method for classifying examples that are not covered by any of the rules in an (unordered) hypothesis. The method, Rule Stretching, is applied after the hypothesis has been induced, during the classi cation phase. Rule Stretching works by generalising the rules ....

W. Van Laer, L. De Raedt, and S. Dzeroski. On multi-class problems and discretization in inductive logic programming. In Proceedings of the 10th International Symposium on Methodologies for Intelligent Systems. Springer-Verlag, 1997.

....[14] In this case, the rst rule that covers the example assigns its label to it. The learning process consists of generating a rule for the class with most uncovered examples and iterate until all the examples are covered. Figure 1 shows an example of such a decision list. The second approach [19] is to learn di erent theories independently for each class, apply all the rules to the new example and if there is a con ict, solve it by considering the distribution of training examples covered by the rules. 4 Experimental Validation Datasets for classi cation of time series are not easy to ....

W. Van Laer, L. De Raedt, and S. Dzeroski. On multi-class problems and discretization in inductive logic programming. In 10 International Symposium on Methodologies for Intelligent Systems (ISMIS97). Springer, 1997.

....reference examples or positive and negative) Since the same example can be selected in several iterations, the calculated distances are saved. When there are more than two classes, it is necessary to learn a theory for each class. The question is how to apply these theories to a new example [30]. We have employed two approaches for dealing with multiclass problems. The first one is the use of ordered rules, also named decision lists [23, 19] In this case, the first rule that covers the example assigns its label to it. The learning process consists of generating a rule for the class with ....

....case, the first rule that covers the example assigns its label to it. The learning process consists of generating a rule for the class with most uncovered examples and iterate until all the examples are covered. Figure 1 shows an example of such a decision list. The second approach, proposed by [30] and based on [10] is to learn di#erent theories independently for each class, apply all the rules to the new example and if there is a conflict, solve it by considering the distribution of training examples covered by the rules. 4 Experimental Validation 4.1 Datasets Description The ....

Wim Van Laer, Luc De Raedt, and Saso Dzeroski. On multi-class problems and discretization in inductive logic programming. In Zbigniew W. Ras and Andrzej Skowron, editors, 10 International Symposium on Methodologies for Intelligent Systems (ISMIS97), volume 1325 of Lecture Notes in Artificial Intelligence, pages 277--286. Springer-Verlag, 1997.

....ations. This resulted in the original ICL system as described in [De Raedt and Van Laer, 1995] At that point, ICL was the acronym of Inductive Constraint Logic, or also I see Logic. In the following years, ICL has undergone many changes and several ex tensions have been added (see, e.g. in [Van Laer et al. 1997] and Section 4.7) One of them was the ability to learn a DNF based hypothesis instead of a set of first order clauses. A DNF rule is easier to understand and it is more intuitive for the user. It turned out that both representations and algorithms are each others dual. At that point, we renamed ....

....issue and have a look at the efficient execution of ILP systems. In that same chapter, we introduce the Extended Input Interface which allows to connect to a database through an ODBC connection. Bibliographical note Some parts of Section 5.2 and Section 5. 3 are based on [Van Laer et al. 1996b; Van Laer et al. 1997] 5.2 Handling numerical data Many symbolic learning systems have difficulties when faced with numerical data. These problems affect the efficiency of learning and the accuracy of the learned theory. 93 CHAPTER 5. EXTENSIONS AND EXTRA FEATURES For instance, let us look back at CN2 as ....

[Article contains additional citation context not shown here]

W. Van Laer, L. De Raedt, and S. D2eroski. On multi- class problems and discretization in inductive logic programming. In Zbigniew W. Ras and Andrzej Skowron, editors, Proceedings of the Tenth International Symposium on Methodologies for Intelligent Systems, volume 1325 of Lecture Notes in Artificial Intelligence, pages 277-286. Springer-Verlag, 1997.

....proposes a high level concept specification language called DLAB in which the hypothesis language syntax can be defined. DLAB grammars are preprocessed in order to generate candidate hypotheses from the most general to the specific ones (under # subsumption) ICL enables also multi class learning [7]. The idea beyond multi class learning is simple: when learning one particular class consider as positive only those examples belonging to this class and as negative all the examples belonging to the remaining classes. This is an attractive option in our case as we want to discover definitions ....

W. Van Laer, L. De Raedt, and S. Dzeroski. On multi-class problems and discretization in inductive logic programming. In Proc. of ISMIS97, LNAI, vol.1325. Springer-Verlag, 1997.

.... has some number handling capabilities (using discretization) it can learn multi classes in case there are more than 2 classes (we use a similar strategy as in CN2, but have also a Bayesian method included) it has some built in evaluation methods (e.g. 10 fold cross validation) see also [26]) Several settings can be used to tune the heuristics and a non graphical interface is integrated that makes the use of ICL fairly easy. 3.2 Tilde: Top down induction of logical decision trees Tilde is a member of the popular top down induction of decision tree family of algorithms. However, ....

.... (Q can be any predicate containing the variable V and is used to generate values for V that occur in the examples) The predicate constants returns the list of all possible values for V , while discretize returns a small subset of all numeric values by using a discretization algorithm (details in [26]) The returned list is used as the third argument in the corresponding dlab variable. For the registration application, we would like to classify participants (the examples) into people who attend the party and people who don t. So in the file registration.s we set classes( party(yes) ....

W. Van Laer, L. De Raedt, and S. Dzeroski. On multi-class problems and discretization in inductive logic programming. In Proceedings of the Tenth International Symposium on Methodologies for Intelligent Systems, pages 277--286. SpringerVerlag, 1997.

....other ILP systems can be incorporated. As such, results from ILP can be reused by propositional learners, so both communities can learn from each other. 18 Step 9: Add interesting extra features. The system ICL has many extensions optimizations w.r.t. the basic system described up to now (see [69, 68] for more details) To handle numerical data, we upgraded the discretization method of Fayyad and Irani (see [36, 29] towards ICL. To handle multiple classes, we extended the CN2 method with a Bayes approach (inspired by [54] This result can be integrated in CN2 without any problem ....

....allows tests on the charge of an atom, background 3 adds 2 specific measures w.r.t. the molecule (e.g. log P and ffl LUMO ) and background 4 consists of descriptions of higher level structures that appear in the molecule (like aromatic rings) Experiments with ICL on this data set can be found in [68]. Results with ICL version 4.2 are given in Table 8. We manually discretized the numerical values (i.e. log P , ffl LUMO and the Charge of the atoms) It seems that the multi class theory is always better than the seperate (DNF) theory for each class. This is not so surprising as the multi class ....

[Article contains additional citation context not shown here]

W. Van Laer, L. De Raedt, and S. Dzeroski. On multi-class problems and discretization in inductive logic programming. In Proceedings of the Tenth International Symposium on Methodologies for Intelligent Systems, pages 277--286. Springer-Verlag, 1997.

No context found.

Van Laer, W., De Raedt, L., D zeroski, S.: On Multi-Class Problems and Discretization in Inductive Logic Programming, in: Proceedings of the 10th International Symposium on Methodologies for Intelligent Systems (ISMIS97) (Z. Ra s, A. Skowron, Eds.), vol. 1325 of LNAI, Springer-Verlag, 1997, 277--286.

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