G. James. Majority vote classi ers: theory and applications. PhD thesis, Department of Statistics - Stanford University, Stanford, CA, 1998.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... classi er: James demonstrated that asymptotically, as the the number of dichotomizers approaches in nity, the ECOC classi er will become Bayes consistent (i.e. it always classi es to the Bayes class when the base learner is the Bayes classi er) provided that a random coding matrix is used [27]. In the same perspective Berger showed that randomly selected decomposition matrices are likely to have pairwise well separate codewords, that is high error recovering capabilities [7] Variants of the original ECOC algorithms have been proposed, as circular ECOC [21] in order to reduce the ....

.... to maximize error recovering capabilities through the maximization of the minimum distance between each couple of codewords [32, 36] Several methods for generating ECOC codes have been proposed: exhaustive codes, randomized hill climbing [17] Hadamard and BCH codes [8, 46] and random codes [27], but open problems are still the joint maximization of distances between rows and columns of the decomposition matrix. 2.2 Reconstruction and decoding After the training of the dichotomizers f i , their outputs are used to reconstruct the polychotomy in order to determine the class C i 2 fC 1 ....

[Article contains additional citation context not shown here]

G. James. Majority vote classi ers: theory and applications. PhD thesis, Department of Statistics - Stanford University, Stanford, CA, 1998.

....been reported for ECOC, there is some discussion as to why it works well. A long random code appears to perform as well or better than a code designed for its error correcting properties [9] Attempts have been made to develop a theory for ensemble classi ers in terms of bias variance and margin [8], but so far these ideas have not provided a convincing explaination for ECOC. A practical approach to determining source of e ectiveness of ECOC is to look at variants of the ECOC strategy to see how they perform. This is also useful if we want to extend ECOC to deal with applications for which ....

....with applications for which it would be desirable to understand ECOC features as estimation measures. In this paper we look at an alternative ECOC combination strategy based on Least Squares (LS ECOC) which was investigated in [11] and extended by incorporating ridged regression when b is small [8]. Recovering individual class probabilities from super class probabilities is easily accomplished by matrix inversion when the individual probability estimates are exact and columns of ECOC matrix are arranged in one per class structure. In practice, estimates are not perfect and a natural ....

[Article contains additional citation context not shown here]

G. James. Majority Vote Classiers: Theory and Applications. PhD thesis, Dept. of Statistics, Univ. of Stanford, May 1998. http://www-stat.stanford.edu/ gareth/.

....the selection of a particular learning machine and the e ectiveness of the ECOC decomposition methods. Several methods for generating ECOC codes have been proposed: exhaustive codes, randomized hill climbing [15] 10 Hadamard and BCH codes [9, 29] borrowed from coding theory, and random codes [19], but open problems are still the joint maximization of distances between rows and columns of the decomposition matrix. Another open problem consists in designing codes for a given multiclass problem. An interesting greedy approach is proposed in [25] and a method based on soft weight sharing to ....

G. James. Majority vote classiers: theory and applications. PhD thesis, Department of Statistics - Stanford University, Stanford, CA, 1998.

....into account substantially di erent learning tasks. Conversely, ensemble of learning machines [8, 17] can achieve their best performances if their base components are accurate and diverse [15] Kuncheva and Whitaker [20] have shown that the dependency between classi ers in majority vote ensembles [21, 18] is related to their classi cation accuracy, and Masulli and Valentini [25] have shown that e ectiveness of ECOC ensemble methods [10] depends on the accuracy and independence among the base dichotomic classi ers. The analysis of the dependence among output errors of di erent learning machines ....

G. James. Majority vote classiers: theory and applications. PhD thesis, Department of Statistics - Stanford University, Stanford, CA, 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