Cohen W.W. Learning with Set-valued Features. In Proceedings of the Thirteenth National Conference on Arti cial Intelligence, 1996.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....feature is connected to exactly another one and the class. Forcing the creation of a treestructured network and incorporating in the model only pair wise 1 Note the similarity of this representation with the one used by RIPPER [6] where it is called learning with set valued features (see also [4]) dependencies allows the relaxing of the strong independence assumptions of NB and at the same time prevents the creation of complex Bayesian Network structures that are computationally more expensive often without paying in terms of classification accuracy. In the learning phase TAN first ....

William Cohen, "Learning with set-valued features", Proc. of the 13 th National Conference on Artificial Intelligence (AAAI-96), 1996.

....have found ways to send messages with forged headers. For example, the sender address can be made the same as the receiver address. A more general and effective approach is obviously needed. In [2] Cohen presented an approach to e mail classification in which a learning program, called RIPPER [1, 3], is used to obtain a set of keyword spotting rules. If all the keywords in a rule are found in a message, the conclusion in the rule is drawn. For example, RIPPER created the following set of rules to recognize talk announcements: A message is a talk announcement if it contains: ffl talk and ....

William W. Cohen. Learning with set-valued features. In Proceedings of AAAI-96, 1996.

.... WHIRL to the accuracy of several other inductive learning methods, including Yang and Chute s method (in the table, these results are labeled K NN) 1 NN, using cosine distance as the distance function for instances; C4.5 [37] using a binary representation; and RIPPER [5] using set valued features [6]. We used r = 30 in WHIRL s r materialization 21 step, and K = 30 for Yang and Chute s method. Minimal feature selection was used for C4.5 (only terms appearing in less than three examples were discarded) and no feature selection was used for the remaining learning methods. Standard experimental ....

William W. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Arti cial Intelligence, Portland, Oregon, 1996.

.... Table 11: Models produced by FOIL and C4.5 22 learning systems have been developed that learn concepts involving number restrictions, a closely related concept [ Frazier and Pitt, 1994; Cohen and Hirsh, 1994b ] A boolean version of the indirect ILP approach has been described [ Cohen, 1996 ] but numeric features have not been previously investigated in this context. Morasca and Ruhe [ 1997 ] have explored the use of logistic regression and rough sets to discover correlations between reliability data and a set of di#erent possibly pertitent measures. Their approach is more aligned ....

William W. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

....tests allowed for a real valued feature are of the form f or f , where f is a feature and is a constant number, and the primitive tests allowed for a symbolic feature are of the form f = a i , where f is a feature and a i is a possible value for f . RIPPER also allows set valued features [3]. The value of a set valued feature is a set of atomic symbols, and tests on set valued features are of the form a i 2 f , where f is the name of feature and a i is a possible value (e.g. ul2ancestorTagNames) For two class problems of this sort, RIPPER uses a number of heuristics to build a ....

William W. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

....Wiener et al. 1995; Cohen, 1995b] In this paper, we will investigate the performance of two recently implemented machine learning algorithms on a number of large text categorization problems. The two algorithms considered are set valued Ripper, a recent rule learning algorithm [Cohen, 1995a; Cohen, 1996b] and sleeping experts, a new on line learning method [Freund et al. 1997] These algorithms share several features that make them attractive for large text categorization problems. First, both algorithms are efficient on large, noisy corpora, running in linear or nearly linear time. Second, ....

....a single attribute, having as its value the set of words that appear in the document. The primitive tests on a set valued attribute a (i.e. the tests which are allowed in rules) are of the form w i 2 a . The implementation of this extension is explained briefly below, and more detail elsewhere [Cohen, 1996b] When constructing a rule, Ripper must find a single test that maximizes information gain; most of Ripper s run time is spent in this operation. For set valued attributes that can be done in two steps. First, Ripper iterates over the set of examples S that are covered by the current rule, ....

[Article contains additional citation context not shown here]

William W. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

....Wiener et al. 1995; Cohen, 1995b] In this paper, we will investigate the performance of two recently implemented machine learning algorithms on a number of large text categorization problems. The two algorithms considered are set valued Ripper, a recent rule learning algorithm [Cohen, 1995a; Cohen, 1996] and sleeping experts, a new on line learning method. These algorithms share several features that make them attractive for large text categorization problems. First, both algorithms are efficient on large, noisy corpora, running in linear or nearly linear time. Second, both algorithms use what ....

....finds the test that maximizes information gain for a set of examples S efficiently, making only a single pass over S for each attribute. All symbols w i that appear as elements of attribute a for some training example are considered by Ripper. This extension is described in more detail elsewhere [Cohen, 1996]. In our experiments, a document is generally represented with a single set valued feature, the value of which is the set of all words appearing in the document. In the implementation, the set can include multiple occurrences of an element (extra occurrences are simply ignored) so one can ....

William W. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

....Wiener et al. 1995; Cohen, 1995b] In this paper, we will investigate the performance of two recently implemented machine learning algorithms on a number of large text categorization problems. The two algorithms considered are set valued Ripper, a recent rule learning algorithm [Cohen, 1995a; Cohen, 1996] and a new on line learning method, modified for text categorization, sleeping experts. These algorithms share several features that make them attractive for large text categorization problems. First, both algorithms are efficient on large, noisy corpora, running in linear or nearly linear ....

....ignored) so one can represent a document with the list of words that appear in it. Optionally, Ripper can also include tests of the form w i 62 feature in its rules. Although extending Ripper to find and use such tests is a simple matter, there are some reasons for being wary of using them [Cohen, 1996]. In this paper we will typically present results for two versions of Ripper one with negative word tests allowed, and one with negative word tests forbidden. Unless stated otherwise, Ripper will be used with negative word tests forbidden. A second extension to Ripper motivated by text ....

William W. Cohen. Learning with set-valued features. Submitted to AAAI-96, 1996.

No context found.

WilliamW. Cohen. Learning with set-valued features. Submitted to AAAI-96, 1996.

.... compared the generalization accuracy of WHIRL to the accuracy of several other inductive learning methods, including Yang and Chute s method (K NN) 1 NN, using cosine distance as the distance function for instances; C4.5 [30] using a binary representation; and Ripper [4] using set valued features [5]. We used r = 30 in WHIRL s r materialization step, and K = 30 for Yang and Chute s method. Minimal feature selection was used for C4.5 (terms appearing in less than three examples were discarded) and no feature selection was used for the remaining learning methods. Standard experimental ....

William W. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

No context found.

WilliamW. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

No context found.

WilliamW. Cohen. Learning with set-valued features. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, Portland, Oregon, 1996.

No context found.

Cohen W.W. Learning with Set-valued Features. In Proceedings of the Thirteenth National Conference on Arti cial Intelligence, 1996.

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