H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....(or a Prolog knowledge base) described by a set of facts. Logical decision trees can introduce variables in the tests that they use and reference these variables in tests at lower nodes. RRL uses logical decision trees as implemented in the programs TILDE [2] for classi cation) and TILDE RT [1] (for regression) 3.3 TG algorithm Classi cation and regression trees are typically induced using a divide and conquer algorithm, called top down induction of decision trees (TDIDT) The reader can consult [9] or [2] for more information. However, since reinforcement learning is an incremental ....

Blockeel, H., De Raedt, L., and Ramon, J. Top-down Induction of Clustering Trees. In Proc. 15th Intl. Conf. on Machine Learning, 1998.

....from one object to another. The means to describe groups of such objects in terms of occurrence of a certain substructure are simply not available in propositional (attribute value) decision trees. This limitation of propositional decision trees has been overcome by an algorithm called Tilde [2, 3]. The approach taken there is to go from propositional to first order representations of objects, and to use first order logic to represent decisions in the tree. The trees that this algorithm produces are called first order logical decision trees. Because of the richer formalism that is used, ....

....between search process and data processing can be made. This enables the data processing part (usually the main computational bottleneck) to be implemented on a scalable server. Only two of these requirements are met by algorithms for inducing first order logical decision trees, as described in [2, 3]. Specifically the items 1. and 3. are addressed by this approach, but little attention has been giving to efficient implementations. The concepts addressed in item 3. are partially solved by representing the whole decision tree as a decision list in Prolog that depends heavily on the order of ....

[Article contains additional citation context not shown here]

Blockeel, H., De Raedt, L., Ramon, J. Top-down induction of clustering trees, In Proceedings of the 15th International Conference on Machine Learning (ICML'98), 55-63, 1998, http://www.cs.kuleuven.ac.be/~ml/PS/ML98-56.ps

.... or objects are repeatedly merged. However, Isaac differs from statistical agglomerative approaches in several ways. First, it is is intended to allow users to guide the construction of the cluster hierarchy which better suits their needs. The user can use the NG parameter, which is in the [0,1] range, to specify both the number of levels and their generality in the hierarchy. As the NG value increases, the system creates more general partitions with few concepts. Lower NG values instruct the system to build more specific partitions. The user can interact with the system experimenting ....

....that, in turn, provides a confirmation on the validity of this knowledge. Another paradigm that automatically integrates background knowledge into the learning process is Inductive Logic Programming (ILP) although there is a small body of research in clustering in this area. A recent exception is [1], although its evaluation is very limited with respect to the typical UCI data sets used for conceptual clustering. Moreover, agglomerative methods like Isaac are common in hierarchical clustering. Since our method does not depend on any particular feature of the system, we think that it could be ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the Fifteenth International Conference on Machine Learning. Morgan Kaufmann, 1998.

No context found.

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998.

....as players usually face each other at the playing board and hence know their opponent, it is an obvious evaluation criterion to validate our approach. In practice one will usually want to predict particular features of the play of the opponent and for this one can use the same method. Tilde [3, 4] is a learning system that builds logical decision trees. It is a generalisation of the propositional learners C4.5 and TDIDT. Logical decision trees are extensions of decision trees to rst order logic. Using rst order logic is a more expressive language than the more frequently used ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998. http://www.cs.kuleuven.ac.be/~ml/PS/ML98-56.ps.

....prediction. The standard TDIDT algorithm can be used: as a heuristic for selecting tests to include in the tree, we use the minimization of intra cluster variance (and maximization of inter cluster variance) in the created clustering. A detailed description of the algorithm can be found in [4]. An implementation is publicly available in the rst order learner Tilde that is included in the ACE tool [5] however for this paper we have used Clus, a downgrade of Tilde that works only on propositional data. 3.2 Ranking via Predicting Errors The instance based approaches to ranking predict ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proc. of the Fifteenth International Conference on Machine Learning, pages 55-63. Morgan Kaufmann, 1998.

....This would not be guaranteed if independent models were learnt for all di erent classes. Now that the problem is clearly de ned, the question is how to construct an algorithm that learns predictive models for this setting. In this paper we follow the predictive clustering approach presented in [4], for which it has been argued that it provides a very general approach to predictive modelling. 4 Predictive Clustering Trees A variety of algorithms for predictive modeling exists. Among the better known are algorithms that induce decision trees [6, 18] Compared to other well known techniques ....

....can build decision trees for multi target prediction. Similarly, if a distance on structured target values is de ned, we can build decision trees for prediction of structural target variables. The methodology has been used successfully for a variety of applications such as conceptual clustering [4], simultaneous prediction of multiple parameters [5] and ranking tasks [23] The algorithm for inducing such trees is essentially a standard TDIDT (Topdown induction of decision trees) algorithm such as ID3 [17] The general idea is to recursively partition a set of data into clusters in such a ....

[Article contains additional citation context not shown here]

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998.

....The standard TDIDT algorithm can be used: as a heuristic for selecting tests to include in the tree, we use the minimization of intra cluster variance (and maximization of inter cluster variance) in the created clustering. A detailed description of the algorithm (called TIC) can be found in [3]. We used the implementation of TIC as available in the first order learner TILDE that is included in the ACE tool [4] This implementation allows for relational tests to be used in the nodes of predictive clustering trees through the use of declarative bias. 3.2 Ranking via Predicting Errors ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998. http://www.cs.kuleuven.ac.- be/ml/PS/ML98-56.ps.

....is used. TILDE has a separate mode, called regression mode, to predict real numbers instead of classes. In that case, a leaf is assigned the mean of the target value and the best test in a node is the test that minimizes the variance in its branches. This setting is described in detail in [2]. link(G,F) group(F,black) liberty cnt(F,LC) LC =2 normal move normal move illegal move suicide capturing move liberty(G,L) link(G,E) group(E,white) liberty cnt(E,1) illegal move notempty no yes yes no yes yes no no move(black, X,Y) group on pos( X,Y) 0,0) G) group(G,empty) Fig. 7. ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998. http://www.cs.kuleuven.ac.be/ml/PS/ML98-56.ps.

....measures. The problem we study is the following: given some set X and a metric d on X, how can we extend d into a metric on the set of all ( nite) subsets of X. Distances between composed objects and between sets of objects have applications in many domains such as cluster analysis (e.g. TIC [2], KBG [1] computational geometry [8] machine learning (e.g. 9, ch.4] RIBL [6] Existing proposals for measures between point sets all have some problems: some are trivial and not very well suited for applications (e.g. the Hausdor metric) others do not satisfy all the properties ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998. http://www.cs.kuleuven.ac.be/~ml/- PS/ML98-56.ps.

....action in a given state, i.e. given a stateaction pair, the P function describes whether this pair is a part of an optimal policy. The P function is represented by a logical decision tree, called the P tree. More information on logical decision trees (classi cation and regression) can be found in [1, 2]. 2.1 The Original RRL Implementation Figure 1 presents the original RRL algorithm. The logical regression tree that represents the Q function in RRL is built starting from a knowledge base which holds correct examples of state, action and Q function value triplets. To generate the examples for ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998. http://www.cs.kuleuven.ac.be/~ml/PS/ML98-56.ps.

.... [1] is an ILP system that induces so called rst order logical decision trees (FOLDT s) Such trees are the rst order equivalent of classical decision trees [12] TILDE can induce classi4 cation and regression trees, but also clustering trees (which allow to predict multiple values at once) see [2] for details. Examples of these di erent kinds of trees will be given in this paper. While for the experiments described in this paper it is not strictly necessary to have an ILP system, as the data can be transformed into single table format, ILP systems typically o er a lot of exibility that ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998. http:- //www.cs.kuleuven.ac.be/~ml/PS/ML98-56.ps.

....is nominal, respectively numerical. For rule based systems, each rule body describes one cluster; for tree based systems the leaves of the tree (in some approaches also the internal nodes) are clusters described by the tests in the tree. In our experiments we used the decision tree learner TILDE [2, 3]. TILDE is an ILP system 4 that induces so called first order logical decision trees (FOLDT s) 4 Inductive logic programming (ILP) is a subfield of machine learning where first order logic is used to represent data and hypotheses. First order logic is more expressive than the attribute value ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. Proc. 15th Int'l Conf. on Machine Learning, pages 55--63, 1998.

....measures. The problem we study is the following: given some set X and a metric d on X , how can we extend d into a metric on the set of all (finite) subsets of X . Distances between composed objects and between sets of objects have applications in many domains such as cluster analysis (e.g. TIC [2], KBG [1] computational geometry [8] machine learning (e.g. 9, ch.4] RIBL [6] Existing proposals for measures between point sets all have some problems: some are trivial and not very well suited for applications (e.g. the Hausdorff metric) others do not satisfy all the properties ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998. http://www.cs.kuleuven.ac.be/~ml/- PS/ML98-56.ps.

....or a distance function. Both tools are in the propositional case a function of values of attributes and are therefore dicult to upgrade to rst order logic. Systems that cluster examples represented in rst order logic exist, but often still need some propositional information. e.g. the system TIC [2] builds clustering trees but uses a distance measure between examples and or clusters. While distance measures between rst order objects exist [12] 7] they are often quite ad hoc, computational expensive and not directed to the formation of conceptual clusters. In this paper we present an ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998. http://www.cs.kuleuven.ac.be/~ml/PS/- ML98-56.ps.

....These systems represent objects as models or clauses. This paper develops a general framework for distances between such objects and reports a preliminary evaluation. Keywords : Machine learning, distances, first order logic. 1 Introduction In learning systems based on clustering (e.g. TIC [3], KBG [1] and in instance based learning (e.g. 10, ch.4] RIBL [8] a measure of the distance between objects is an essential component. Good measures exist for distances between objects in an attribute value representation (see e.g. 10, ch. 4] Recently there is a growing interest in using ....

....classification systems (which make use of information that assigns a class to each example in the training set during the building of the decision tree) TIC is a clustering system based on TILDE. Clustering does not make use of class information during the building of the decision tree (See also [3] for more information on evaluating clustering trees via prediction) TIC uses a distance measure for choosing the best tests. Unfortunately, until now only euclidian distances (on the propositional part of the data) could be used. Using a first order measure much better results can be reached. We ....

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998. http://www.cs.kuleuven.ac.be/~ml/PS/ML98-56.ps.

No context found.

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In Proceedings of the 15th International Conference on Machine Learning, pages 55--63, 1998.

No context found.

Blockeel, H., De Raedt, L., and Ramon, J. (1998): Top-down induction of clustering trees. Proceedings of the 15th International Conference on Machine Learning, pages 55--63, Morgan Kaufmann.

No context found.

Hendrik Blockeel, Luc De Raedt, and Jan Ramon. Top-down induction of clustering trees. In J. Shavlik, editor, 15th ICML, pages 55--63. Morgan Kaufmann, 1998.

No context found.

Hendrik Blockeel, Luc De Raedt, and Jan Ramon. Top-down induction of clustering trees. In J. Shavlik, editor, Proceedings of the 15th International Conference on Machine Learning, pages 55-63. Morgan Kaufmann, 1998.

No context found.

H. Blockeel, L. De Raedt, and J. Ramon. Top-down induction of clustering trees. In ICML-98, Morgan Kaufmann, 1998.

No context found.

H. Blockeel, L. DeRaedt, and J. Ramon. Top-down induction of clustering trees. In J. Shavlik, editor, Proceedings of the 15th International Conference on Machine Learning, pages 55-63, 1998.

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