C. Cardie. Improving minority class prediction using casespecific feature weights. In Proceedings of the Fourteenth International Conference on Machine Learning, pages 5765, Morgan Kaufmann, 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....accuracy for NN algorithms appears to be relevant. The NN classification procedure is straightforward: Given a set of classified points in an input space, a new point is assigned to the known class of the nearest one with respect to a metric defined on the input space. Many researchers [SF80, SF81, SW86, CMSW92, AG90, AG92, HT95, Fri94, RA95, RA98, CH97, WMA97] focused their attention on the use of local metrics, i.e. metrics that vary depending on the position of the points in the input space, in order to outperform systems based on global metrics. The claim appears to be controversial. On one hand the local metrics generate classifiers that are more ....

C. Cardie and N. Howe. Improving minority class prediction using case-specific feature weight. In Proceedings of the Fourteenthn International Conference on Machine Learning, pages 57--65. Morgan Kaufmann, 1997.

.... make use of conditional probabilities (Crecy et al. 37] class projections (Stanfill and Waltz [38] Howe and Cardie [39] mutualinformation (Wettschereck and Dietterich [40] in a nearest hyperrectangle approach) or information gain (van den Bosch and Daelemans [41] Cardie and Howe [42] first build a decision tree to select features and then weight each feature according to its information gain) VI. Summary and future work Once the Feature Weighting problem for nearest neighbor classifier has been stated as a search problem, GAs, due to their attractive randomized and ....

C. Cardie, N. Howe, Improving Minority Class Prediction Using Case-Specific Feature Weights, Proceedings of the Fourteenth International Conference on Machine Learning, Nashville, USA, 1997, pp. 57-65.

....Rate False Positive Rate Figure 1: An idealized ROC curve. predictions on the majority class. As a result, most errors will be on the more expensive minority class. There are several ways to cope with this problem. For example, we can alter the distribution of the training examples (e.g. Cardie Howe, 1997; Freund Schapire, 1996; Kubat Matwin, 1997) or we can change the way the learning methods treat instances from classes with different error costs (e.g. Lewis Catlett, 1994; Maloof, Langley, Sage, Binford, 1997; Bradley, 1997) If we have cost sensitive learning methods and a cost ....

....Task Revisited To demonstrate the value of ROC analysis, we present new results for the recidivism prediction task using the 1978 data set. The distribution of the training examples was 27.5 recidivist and 72.5 non recidivist, which is not skewed as severely as other reported data sets (e.g. Cardie Howe, 1997; Maloof et al. 1997) but it was enough to adversely impact the performance of all of the machine learning methods, most notably the perceptron algorithm. For this problem, we have only an informal notion of error costs: Mistakes on the positive (i.e. the minority) class are more expensive ....

Cardie, C., & Howe, N. (1997). Improving minority class prediction using case-specific feature weights. In Proceedings of the Fourteenth International Conference on Machine Learning (pp. 57--65). San Francisco, CA: Morgan Kaufmann.

....is straightforward: given a set of classified examples, which are described as points in an input space, a new unclassified example is assigned to the known class of the nearest example. The nearest relation is computed using a (similarity) metric defined on the input space. Many researchers [21 23, 11, 1, 2, 14, 13, 18, 19, 7, 25] focused their attention on the use of local metrics, i.e. metrics that vary depending on the position of the points in the input space. Conversely, more traditional global metrics assume that similarity evaluation should be independent from the area of the input space the cases to be compared are ....

C. Cardie and N. Howe. Improving minority class prediction using case-specific feature weight. In Proceedings of the Fourteenth International Conference on Machine Learning, pages 57--65. Morgan Kaufmann Publishers, 1997.

....the real world. Kubat and Matwin (1997) acknowledged the performance degradation effects of skewed class distribution and investigated techniques for removing unnecessary instances from the majority class. Instances that are in the borderline region, noisy, or redundant are candidates for removal. Cardie and Howie (1997) stated that skewed class distributions are the norm for learning problems in natural language processing (NLP) In a case based learning framework, they studied techniques to extract relevant features from previously built decision trees and customize local feature weights for each case ....

Cardie, C., and Howe, N. 1997. Improving minority class prediction using case-specific feature weights. In Proc. 14th Intl. Conf. Mach. Learning, 57--65.

....Kubat and Matwin [14] acknowledged the performance degradation effects of skewed class distributions and investigated techniques for removing unnecessary instances from the majority class. Instances that are in the borderline region, noisy, or redundant are candidates for removal. Cardie and Howie [3] stated that skewed class distributions are the norm for learning problems in natural language processing (NLP) In a case based learning framework, they studied techniques to extract relevant features from previously built decision trees and customize local feature weights for each case ....

C. Cardie and N. Howe. Improving minority class prediction using case-specific feature weights. In Proc. 14th Intl. Conf. Mach. Learning, pages 57--65, 1997.

....thresholds) and IB2 is a simplification of CNN (Hart, 1968) 2 Since these early efforts, lazy algorithms have undergone dramatic design enhancements. Researchers have studied several lazy learning topics, including ffl case selection (Zhang et al. 1997) ffl concept distribution skew (Cardie Howe, 1997), ffl concept shift (Salganicoff, 1997) ffl cost sensitive learning (Tan Schlimmer, 1990; Turney, 1993) ffl discretization (Ting, 1997; Wilson Martinez, 1997a) 1 Several synonyms have been used to describe these algorithms. For example, these include ....

....elsewhere. For example, see (Kurtzberg, 1987) ffl feature selection and weighting (Kelly Davis, 1991; Cain et al. 1991; Cardie, 1993; Skalak, 1994; Ricci Avesani, 1995; Kohavi et al. 1997; Domingos, 1997; Ling Wang, 1997; Maron Moore, 1997; Wettschereck et al. 1997; Howe Cardie, 1997), ffl information theory (Lee, 1994; Cleary Trigg, 1995; Wettschereck Dietterich, 1995) ffl noise (Stanfill, 1987; Aha Kibler, 1989; Aha et al. 1991; Ting, 1997) ffl parallel implementations (Stanfill Waltz, 1986) ffl preference learning (Branting Broos, 1994) ffl speedup ....

Cardie, C., & Howe, N. (1997). Improving minority-class prediction using case-specific feature weights. Proceedings of the Fourteenth International Conference on Machine Learning (pp.

....all values for f . PCF assigns high weights to features that have high correlations with the given class. IB4 (Aha, 1992) mentioned previously, also infers class specific weights. It uses a different weight updating rate depending on class frequency. More recently, Howe and Cardie [ Howe and Cardie, 1997)] introduced CDW, a class specific weighting algorithm for symbolic features that computes weights using w(f; c j ) X v2f jr(f; v; fxjx2c j g) Gamma r(f; v; fxjx62c j g)j (1.14) where r(f; v; X 0 ) is, for a training subset X , P (x 0 f = vjx 0 2X 0 ) CDW sets a feature s weight ....

....samples fall outside of a predetermined range around q. After discarding these samples, the new most relevant feature is determined, and the process is repeated until only k training samples remain. Scythe performed well in comparison to k NN on several classification tasks. Cardie and Howe s [(Cardie and Howe, 1997)] sample specific weighting approach induces a pruned decision tree from the training set to perform feature selection; features in the path that classify q are selected. Continuous weights are then assigned to these features according to their entropy on the entire training set. Their algorithm ....

Cardie, C. and Howe, N. (1997). Improving minority-class prediction using case-specific feature weights. In Proceedings of the Fourteenth ICML, pages 57--65, Nashville, TN. Morgan Kaufmann.

No context found.

C. Cardie. Improving minority class prediction using casespecific feature weights. In Proceedings of the Fourteenth International Conference on Machine Learning, pages 5765, Morgan Kaufmann, 1997.

No context found.

C. Cardie and N. Howe. 1997. Improving minority class prediction using case-specific feature weights. In D. Fisher, editor, Proceedings of the Fourteenth International Conference on Machine Learning. Morgan Kaufmann.

No context found.

C. Cardie and N. Howe, "Improving minority class prediction using case-specific feature weights," in Proc. ICML, pp. 57-65, 1997.

No context found.

Cardie, C., Howe, N.: Improving minority class prediction using case-specific feature weights, Proc. 14 International Conference on Machine Learning, Morgan Kaufmann, 1997, 57--65.

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