W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proc. Fourth International Workshop on Inductive Logic Programming #ILP-94#, 53754 Sankt Augustin, Germany, 1994. GMD. GMD-Studien Nr. 237. .  Home/Search   Document Details and Download   Summary   Related Articles   Check

....that leaf. If the leaves are coherent with respect to classes, this method would yield relatively high classification accuracy with a minimum of class information available. This is quite similar in spirit to Emde s method for learning from few classified examples, implemented in the COLA system [ Emde, 1994 ] A similar reasoning can be followed for regression, leading to unsupervised regression ; again this may be useful in the case of partially missing information. We conclude that clustering can extend classification and regression towards unsupervised learning. Another extension in the ....

....logmutag all three 100 0.80 0.81 50 0.78 0.79 25 0.72 0.77 10 0.67 0.74 Table 4: Classification accuracies obtained for Mutagenesis with several distance functions, and on several levels of missing information. distances. This experiment is similar in spirits to the ones performed with COLA [ Emde, 1994 ] Table 4 shows the results. As expected, performance degrades less quickly when more information is available, which supports the claim that the use of more than just class information can improve performance in the presence of missing information. 6 CONCLUSIONS AND RELATED WORK We have ....

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pages 51--70, Sankt Augustin, Germany, 1994. Gesellschaft fur Mathematik und Datenverarbeitung MBH.

....is one of the fundamental unsupervised learning tasks, and has been intensively studied for propositional representations, both in statistics and in Machine Learning. For relational, first order representations, there has been a lot less research, but here also, existing work (e.g. 20] 1] [10], 3] has shown the application potential of clustering for tasks where class information is sparse, expensive to obtain, or unavailable. However, up to now, work in ILP has mostly concentrated on the task of conceptual clustering, i.e. restricting cluster formation to clusters with symbolic ....

....measure that selects a single level set of clusters from the induced clustering hierarchy. Results from empirical experiments are reported in section 4. In the related work section, we discuss other first order clustering systems, and in particular the relationship of RDBC to KBG [1] Cola 2 [10] and C0.5 [3] which have used distance functions within conceptual clustering. We conclude with a summary and some pointers to future work. 2 Distance Based Clustering For propositional representations, most clustering algorithms are based on elementary distance properties of the instance space. ....

[Article contains additional citation context not shown here]

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proc. Fourth International Workshop on Inductive Logic Programming (ILP-94), 53754 Sankt Augustin, Germany, 1994. GMD. GMD-Studien Nr. 237. .

....is one of the fundamental unsupervised learning tasks, and has been intensively studied for propositional representations, both in statistics and in Machine Learning. For relational, first order representations, there has been a lot less research, but here also, existing work (e.g. 20] 1] [10], 3] has shown the application potential of clustering for tasks where class information is sparse, expensive to obtain, or unavailable. However, up to now, work in ILP has mostly concentrated on the task of conceptual clustering, i.e. restricting cluster formation to clusters with symbolic ....

....measure that selects a single level set of clusters from the induced clustering hierarchy. Results from empirical experiments are reported in section 4. In the related work section, we discuss other first order clustering systems, and in particular the relationship of RDBC to KBG [1] Cola 2 [10] and C0.5 [3] which have used distance functions within conceptual clustering. We conclude with a summary and some pointers to future work. 2 Distance Based Clustering For propositional representations, most clustering algorithms are based on elementary distance properties of the instance space. ....

[Article contains additional citation context not shown here]

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proc. Fourth International Workshop on Inductive Logic Programming (ILP-94), 53754 Sankt Augustin, Germany, 1994. GMD. GMD-Studien Nr. 237. .

....the missing data estimation problem more difficult. Therefore it is interesting to investigate the trade off between the advantage of using the increased information available in the query batch, and the disadvantage of increased complexity in the search process. Similar work has been reported in [2], where the unclassified vectors were used as background knowledge for a conceptual clustering algorithm. In order to study this problem, we use the probabilistic model family of finite mixtures [3, 7] where the problem domain probability distribution is approximated as a finite, weighted sum of ....

W. Emde. Inductive learning of characteristic concept descriptions from small sets of classified examples. In F. Bergadano and L. De Raedt, editors, Proceedings of the 7th European Conference on Machine Learning (ECML94), pages 103--121, 1994.

....that leaf. If the leaves are coherent with respect to classes, this method would yield relatively high classification accuracy with a minimum of class information available. This is quite similar in spirit to Emde s method for learning from few classified examples, implemented in the COLA system [ Emde, 1994 ] 4.3 Regression The above shows that first order clustering can be used for characterisation of clusters in the data, as well as for classification. An application that has a flavour of both, is predicting numerical values. If clusters are coherent with respect to some numerical attribute of ....

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMDStudien, pages 51--70, Sankt Augustin, Germany, 1994. Gesellschaft fur Mathematik und Datenverarbeitung MBH.

....The similarity measure of Bisson takes as input a net of facts about a part of a domain, such that the similarity between a pair of arguments a b may depend on the similarity of another pair of facts c d. The similarity of c d may in turn be computed from the similarity of a b [ Emde, 1994b ] ffl The similarity measure of Ribl can be regarded as a generalization of similarity measures employed in attribute value instance based learners. This is not true for Bisson s measure, due to a difference in the formula to compute the similarity of literals (17) 3.4 The k Nearest Neighbor ....

W. Emde. Inductive learning of characteristic concept descriptions from small sets of classified examples. In F. Bergadano and L. De Raedt, editors, Machine Learning: ECML-94, European Conference on Machine Learning, 1994, volume 784 of Lecture Notes in Artificial Intelligence, pages 103--121, Berlin, 1994. Springer-Verlag.

....sort person have to be treated as name arguments and arguments of sort age are numbers. 3. 2 Generation of Cases from unstructured Theories Instance based learners, as well as many ILP algorithms such as Golem [ Muggleton and Feng, 1992 ] Clint [ De Raedt and Bruynooghe, 1992 ] and Cola [ Emde, 1994a ] require the construction of cases (or starting clauses) from ground facts that can be derived from a knowledge base. The construction of case descriptions is generally restricted by a syntactic or semantic bias (e.g. restricted by a depth parameter) for reasons of computational complexity ....

.... for First Order Logic The similarity computation for Ribl is a modified version of a measure for first order logic representations proposed by Bisson [1992] A measure similar to that by Bisson [1992] has previously been shown to be useful within a first order conceptual clustering system [ Emde, 1994a ] In contrast to Bisson s measure, our measure can be regarded as a natural extension of similarity measures for attribute value representations. It is designed to be applicable to case descriptions as they are described in the previous section. Therefore, our measure can be used in systems ....

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proc. Fourth International Workshop on Inductive Logic Programming (ILP-94), Arbeitspapiere der GMD, 53754 Sankt Augustin, Germany, 1994. GMD.

....neighbor learning module, and ffl a module for the similarity computation of pairs of cases. 2. 1 Generation of Cases from unstructured Theories Instance based learners as well as many ILP algorithms (e.g. GOLEM [ Muggleton and Feng, 1992 ] CLINT [ De Raedt and Bruynooghe, 1992 ] COLA [ Emde, 1994 ] require the construction of cases ( starting clauses ) from all ground facts that can be derived from a knowledge base. The construction of cases is generally performed by an approximate process (e.g. restricted by a depth parameter) for reasons of computational complexity. In Mobal, a ....

.... using such descriptions, a rule learning system would be able to induce a rule such as: high quality workshop(X) annual fgml workshop(X) organizer(X,Y) organizer(X,Z) Y 6= Z Details about the current case generation module taken from the Cola 2 system are described in [ Emde, 1994 ] 2.2 Similarity Computation in First Order Logic The similarity computation for Ribl is an unweighted version of the measure initially proposed by Bisson (1992) The main idea of Bisson s similarity measure is to compute the similarity of all entities involved in two case descriptions. The ....

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proc. Fourth International Workshop on Inductive Logic Programming (ILP-94), Arbeitspapiere der GMD, 53754 Sankt Augustin, Germany, 1994. GMD.

No context found.

W. Emde. Inductive learning of characteristic concept descriptions. In S. Wrobel, editor, Proc. Fourth International Workshop on Inductive Logic Programming #ILP-94#, 53754 Sankt Augustin, Germany, 1994. GMD. GMD-Studien Nr. 237. .

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