Ron Kohavi. Bottom-up induction of oblivious read-once decision graphs. In European Conference in Machine Learning, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... An excellent review of other related work in PAC learning that uses structural bias and queries is given in [48] Function decomposition is also related to construction of oblivious read once decision graphs (OODG) OODGs are rooted, directed acyclic graphs that can be divided into levels [17]. All nodes at a level test the same attribute, and all edges that originate from one level terminate at the next level. Like with decision trees, OODG leaf nodes represent class values. OODGs can be regarded as a special case of decomposition, where decomposition structures 30 are of the form f ....

....each level equals the number of distinct output values used by corresponding function f i . In fact, decision graphs were found as a good form of representation of examples to be used by decomposition [21, 20, 13] Within machine learning, the use of oblivious decision graphs was studied by Kohavi [17]. Graphs induced by his learning algorithm are consistent with training examples, and for incomplete datasets the core of the algorithm is a graph coloring algorithm similar to the one defined by Perkowski and Uong [34] Of other machine learning approaches that construct concept hierarchies we ....

R. Kohavi. Bottom-up induction of oblivious read-once decision graphs. In F. Bergadano and L. De Raedt, editors, Proc. European Conference on Machine Learning, pages 154--169. Springer-- Verlag, 1994.

....model of computation that takes the form of a directed acyclic graph (DAG) BPs have been well studied in theoretical computer science. In this paper we view them as classiers. In empirical machine learning a similar representation formalism decision graphs has been studied to some extent [14, 10, 11]. BPs are a strict generalization of DTs. Thus, their learning in computational learning frameworks is hard [5, 1] Mansour and McAllester [13] devised a boosting algorithm for BPs. The main advantage obtained by using BPs rather than DTs is that their training error is guaranteed to decline ....

....Length (MML) criterion to determine whether to split a leaf or to join a pair of leaves in the evolving DG. In experiments the algorithm attained the same accuracy level as MML based DT learners and C4.5. DGs were observed to give particularly good results in learning disjunctive concepts. Kohavi [10], originally, proposed constructing DGs in a bottom up manner. Despite some empirical success, the algorithm was not able to cope with numerical attributes and lacked methods for dealing with irrelevant attributes. Subsequently Kohavi and Li [11] presented a method that post processes a DT ....

Kohavi, R.: Bottom-up induction of oblivious read-once decision graphs. In: Bergadano, F., De Raedt, L. (eds.): Machine Learning: Proc. Seventh European Conference. Lecture Notes in Articial Intelligence, Vol. 784, Springer-Verlag, Berlin Heidelberg New York (1994) 154169

....the replication problem. The fragmentation problem has been attacked in different ways: by constructing compound features at every tree node [12, 18] by reducing the number of possible partitions [5, 16] and by using alternative concept representations, e.g. sets of rules [15] decision graphs [6, 11], SE Trees [22] decision lists [12, 21] Nonetheless, no clear solution has emerged. Appeared in the 9th European Conference on Machine Learning (1997) Lecture Notes in Artificial Intelligence, Vol. XXX, Springer Verlag, Heidelberg, pp 312 326. Our analysis of the causes and effects of the ....

Kohavi R.: Bottom-Up Induction of Oblivious Read-Once Decision Graphs. In Proceedings of the European Conference on Machine Learning, (1994) 154--169

....this problem. Use of compound questions has been a prime candidate, however the cost of selecting the best set of features for a test at a node also increases with number of features. A variety of methods have been attempted to address the problem of data sparsity, including use of decision graphs [76, 96], pylons [4] and soft decisions [107] for additional pointers, see Murthy s comprehensive multi disciplinary survey on decision trees [91] however, no clear solution has emerged. 5.2 Multi stage Clustering Our approach to reduce the storage and computational costs for clustering is based on ....

Ron Kohavi. Bottom-up induction of oblivious read-once decision graphs. Proc. European Conf. on Machine Learning, pages 154-169, 1994.

....we consider three classification techniques for predicting the performance of software with implementations available in the MLC [KSD96] software library and the CN2 [CN89] program. The classifiers we select are CN2 [CN89] using rule induction) ID3 [Qui86] using decision trees) and HOODG [Koh94] using graphs) We describe these algorithms in terms of the case study DP software and its performance database. CN2 generates an ordered list of simple, propositional like classification rules such as if then statements where the conditional part consists of a propositional expression ....

....that a simple but not necessarily the simplest tree is found. A simple decision tree can be generated by selecting a metric which minimizes the information needed in the resulting subtrees to differentiate subsets of performance data based on their corresponding DP algorithms. HOODG [Koh94] represents the underlying knowledge as a decision graph. This graph is directed and acyclic with the following properties: a) there are exactly as many nodes in the graph that have out degree zero as the number of DP algorithms, b) the nodes that do not belong in a) are labeled by some ....

[Article contains additional citation context not shown here]

Ron Kohavi. Bottom-up Induction of Oblivious Read-Once Decision Graphs: Strengths and Limitations. In Proc. 12th National Conference on Artificial Intelligence, pages 613--618, 1994.

....performs better than the original Chi2 algorithm. It becomes a completely automatic discretization method. Index Terms Discretization, degree of freedom, X 2 test. I INTRODUCTION MANY algorithms developed in the machine learning community focus on learning in nominal feature spaces [1]. However, many of such algorithms cannot be applied to the real world classification tasks involving continuous features before these features are first discretized. This demands the studies on the discretization methods There are three different axes by which discretization methods can be ....

R. Kohavi, "Bottom-Up Induction of Oblivious Read-Once Decision Graphs: Strengths and Limitation," Proc. 12th Nat'l Conf. Artificial Intelligence, pp. 613-618, 1994.

....(DAG) and it is leveled if the arcs from the nodes at level l go to those at level l 1. An induction algorithm based on boosting [19, 6, 7] for such BPs was recently devised by Mansour and McAllester [13] Related graph formed classiers have, though, been studied in machine learning also before [15, 11, 12]. Empirical experiments with the BP induction algorithm demonstrate that it produces approximately as accurate and as compact unpruned classiers as DT learning algorithms do [3] In theory BPs should obtain the same (training set) accuracy level as DTs with exponentially smaller classiers [13] ....

Kohavi, R. (1994) Bottom-up induction of oblivious read-once decision graphs: strengths and limitations. In Proceedings of the Twelfth National Conference on Articial Intelligence, pp. 613-618. Menlo Park, CA: AAAI Press.

....model of computation that takes the form of a directed acyclic graph (DAG) BPs have been well studied in theoretical computer science. In this paper we view them as classi ers. In empirical machine learning a similar representation formalism decision graphs has been studied to some extent [14, 10, 11]. BPs are a strict generalization of DTs. Thus, their learning in computational learning frameworks is hard [5, 1] Mansour and McAllester [13] devised a boosting algorithm for BPs. The main advantage obtained by using BPs rather than DTs is that their training error is guaranteed to decline ....

....Length (MML) criterion to determine whether to split a leaf or to join a pair of leaves in the evolving DG. In experiments the algorithm attained the same accuracy level as MML based DT learners and C4.5. DGs were observed to give particularly good results in learning disjunctive concepts. Kohavi [10], originally, proposed constructing DGs in a bottom up manner. Despite some empirical success, the algorithm was not able to cope with numerical attributes and lacked methods for dealing with irrelevant attributes. Subsequently Kohavi and Li [11] presented a method that post processes a DT ....

Kohavi, R.: Bottom-up induction of oblivious read-once decision graphs. In: Bergadano, F., De Raedt, L. (eds.): Machine Learning: Proc. Seventh European Conference. Lecture Notes in Articial Intelligence, Vol. 784, Springer-Verlag, Berlin Heidelberg New York (1994) 154169

....contain a compound premise, and interior nodes to 2 contain conclusions which are asserted if the tree cannot be traversed further. Gaines (1991) shows that ripple down rules are a particular case of rules with exceptions that can encode some knowledge structures more compactly. Oliver (1993) and Kohavi (1994) have shown how various forms of decision graphs may be induced and provide a more compact alternative than decision trees. Gaines (1995) generalizes these representations into exception directed acyclic graphs (EDAGs) that support exceptions within a general decision graph, and subsumes trees and ....

Kohavi, R. (1994). Bottom-up induction of oblivious read-once decision graphs: strengths and limitations. AAAI'94: Proceedings of the Twelfth National Conference on Artificial Intelligence. pp.613-618. Menlo Park, California, AAAI Press/MIT Press.

.... will exhibit results from C4.5 (a typical machine learning algorithm) Espresso (a 2 level minimization circuit design tool) and Function Extrapolation by Recomposing Decompositions (FERD) 1 Introduction Only recently have ideas from logic synthesis and machine learning begun to converge [4] [8]. The goals are similar. In KDD and ML, at the core of every problem is how to effectively generalize concepts from data [11] In the circuit design community, the end product is the realization of a minimum complexity circuit with respect to the number of gates, depth, inputs, literals, or ....

Ron Kohavi. Bottom-up induction of oblivious read-once decision graphs. In European Conference on Machine Learning, 1994.

.... An excellent review of other related work in PAC learning that uses structural bias and queries is given in [40] Function decomposition is also related to construction of oblivious read once decision graphs (OODG) OODGs are rooted, directed acyclic graphs that can be divided into levels [16]. All nodes at a level test the same attribute, and all edges that originate from one level terminate at the next level. Like with decision trees, OODG leaf nodes represent class values. OODGs can be regarded as a special case of decomposition, where decomposition structures are of the form f 1 (x ....

....each level equals the number of distinct output values used by corresponding function f i . In fact, decision graphs were found as a good form of representation of examples to be used by decomposition [20, 19, 13] Within machine learning, the use of oblivious decision graphs was studied by Kohavi [16]. Graphs induced by his learning algorithm are consistent with training examples, and for incomplete datasets the core of the algorithm is a graph coloring algorithm similar to the one defined by Perkowski and Uong [33] Of other machine learning approaches that construct concept hierarchies we ....

R. Kohavi. Bottom-up induction of oblivious read-once decision graphs. In F. Bergadano and L. De Raedt, editors, Proc. European Conference on Machine Learning, pages 154--169. Springer-- Verlag, 1994.

....also great wins for successful programs. The first research results that appreciate this synergy and try to link the two worlds of the machine learning community and the design automation community already start to appear: new decision diagram approaches were presented in 1994 by Ron Kohavi [10, 11], and Arlindo Oliveira [15] It is our hope that the participants of this Workshop will also partake in this challenge, develop new theories and software, and will test them on the machine learning benchmarks. It is quite possible, that problems with an unusually high percent of don t cares will ....

R. Kohavi, "Bottom-up Induction of Oblivious Read-Once Decision Diagrams," In European Conference on Machine Learning, 1994.

....one reduces the description length to a larger extent. He reported improvements over the use of decision trees on relatively simple problems, but our experiments using a similar approach failed on more complex test cases because the algorithm tends to perform premature joins on complex problems. Kohavi (1994) proposed an approach that also uses reduced ordered decision graphs. His approach builds an ordered decision graph in a bottom up fashion, starting at the level closest to the terminal nodes. This choice is partially based on the fact that the widest level (i.e. the level with the larger number ....

Kohavi, R. (1994). Bottom-up induction of oblivious read-once decision graphs: Strengths and limitations. In Twelfth National Conference on Artificial Intelligence, Tahoe City, CA, pages 613--618. Morgan Kaufmann.

....when trying to identify common subtrees. We describe a simple but relatively effective algorithm that derives an ordered decision graph with minimal description length from a decision tree built using standard techniques. The approach proposed here is very different from the one proposed in [Koh94] that also used RODGs. His approach as described in the reference, suffers from serious limitations when solving problems with medium to large numbers of attributes although it performs well for small problems. The complexity of the generated decision graph and the quality of the generalization ....

Ron Kohavi. Bottom-up induction of oblivious read-once decision graphs. In European Conference in Machine Learning, 1994.

.... design tool) and Function Extrapolation by Recomposing Decompositions (FERD) 1 Introduction Although we have seen circuit design techniques applied to ML and KDD problems since the 1960 s from Michalski, only recently have ideas from logic synthesis and machine learning begun to converge [5, 10, 8]. The tools from the early days of circuit design have advanced a long way from Karnaugh maps where sophisticated software has been developed for over 25 years. In the ML community, we have seen papers on Presented at the November 1995 IEEE International Conference for Tools with Artificial ....

Ron Kohavi. Bottom-up induction of oblivious readonce decision graphs. In European Conference on Machine Learning, 1994.

....is added to the window and then used to grow the decision tree. This algorithm is based on the idea that it is less profitable to consider the training set, in its entirety, than an appropriately chosen part of it. ffl HOODG This is a greedy hill climbing inducer for building decision graphs [34]. It does this in a bottom up manner. It was originally proposed to overcome the disadvantages of decision trees duplication of subtrees in disjunctive concepts (replication) and partitioning of data into fragments, where a high arity attribute is tested at each node (fragmentation) Thus, it ....

R. Kohavi, "Bottom-up induction of oblivious, read-once decision graphs: Strengths and limitations," in Twelfth National Conference on Artificial Intelligence, 1994, pp. 613--618, ftp://starry.stanford.edu/pub/ronnyk/aaai94.ps.

....simple problems but our experiments using a similar approach failed in more complex test cases because the algorithm tends to perform premature joins in complex problems. Very recently, a new approach was proposed [Kohavi, 1995] that reportedly represents an improvement over a previous algorithm [Kohavi, 1994] by the same author. This last approach is also based on the identification of common subtrees in a decision tree, but this tree is constrained to exhibit the same ordering of tests for all possible paths in the tree. A combination of this approach with some of the techniques introduced here may ....

Kohavi, R. (1994). Bottom-up induction of oblivious read-once decision graphs. In European Conference in Machine Learning.

....rules generalize binary decision trees by allowing a node to contain a compound premise, and interior nodes to contain conclusions. Gaines (1991) shows that ripple down rules are a particular case of rules with exceptions that can encode some knowledge structures more compactly. Oliver (1993) and Kohavi (1994) have shown how various forms of decision graphs may be induced and provide a more compact alternative than decision trees. Gaines (1995) generalizes these representations into exception directed acyclic graphs (EDAGs) that support exceptions within a general decision graph, and subsumes trees and ....

Kohavi, R. 1994. Bottom-up induction of oblivious readonce decision graphs: strengths and limitations.

No context found.

Ron Kohavi. Bottom-up induction of oblivious read-once decision graphs. In European Conference in Machine Learning, 1994.

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