T. Mitchell. Conditions for the equivalence of hierarchical and flat bayesian classifiers. Technical report, Center for Automated Learning and Discovery, Carnegie-Mellon University, 1998.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... categorisers 4 performed no better than a single Bayesian categoriser when applied to categorising the Reuters corpus [17] This empirical result was formalised by Mitchell, who derived a proof that a hierarchical organisation of homogeneous categorisers is equivalent to a single flat categoriser [40]. However, Koller and Sahami also discovered that a hierarchical organisation of heterogeneous Bayesian categorisers, applied to categorising the Reuters corpus, achieved a significant improvement in accuracy. In particular, they argued that since each categoriser faces a small, individual, ....

....documents, the hierarchical organisation of Bayesian categorisers (HOC) achieves an accuracy almost identical to the flat Bayesian categoriser (NB) when there is a dearth of training. This is consistent with Mitchell s proof of the equivalence of hierarchical and flat probabilistic categorisers [40]. However, it can also be seen in figure 6.1 that as training documents become abundant, NB out performs HOC. The difference between NB and HOC may be explained by considering two competing influences that govern the performance of HOC: 1. A document must pass through a number of decision points ....

[Article contains additional citation context not shown here]

Tom M. Mitchell. Conditions for the Equivalence of Hierarchical and flat Bayesian Classifiers. Technical report, Center for Automated Learning and Discovery, Carnegie-Mellon University, February 1998.

.... categorisers 4 performed no better than a single Bayesian categoriser when applied to categorising the Reuters corpus [17] This empirical result was formalised by Mitchell, who derived a proof that a hierarchical organisation of homogeneous categorisers is equivalent to a single flat categoriser [40]. However, Koller and Sahami also discovered that a hierarchical organisation of heterogeneous Bayesian categorisers, applied to categorising the Reuters corpus, achieved a significant improvement in accuracy. In particular, they argued that since each categoriser faces a small, individual, ....

....documents, the hierarchical organisation of Bayesian categorisers (HOC) achieves an accuracy almost identical to the flat Bayesian categoriser (NB) when there is a dearth of training. This is consistent with Mitchell s proof of the equivalence of hierarchical and flat probabilistic categorisers [40]. However, it can also be seen in figure 6.2 that as training documents become abundant, NB out performs HOC. The difference between NB and HOC may be explained by considering two competing influences that govern the performance of HOC: 1. A document must pass through a number of decision points ....

[Article contains additional citation context not shown here]

Tom M. Mitchell. Conditions for the Equivalence of Hierarchical and flat Bayesian Classifiers. Technical report, Center for Automated Learning and Discovery, Carnegie-Mellon University, February 1998.

.... large vocabulary sizes perform best [ Joachims, 1997; Nigam et al. 1998 ] Somewhat surprisingly, it can be shown that a pure form of Pachinko Machine classification using maximum likelihood estimates and a constant vocabulary is in fact equivalent to performing non hierarchical classification [ Mitchell, 1998 ] The remainder of this paper is structured as follows: we explain our probabilistic approach to text classification, and present the use of shrinkage in this context. Then we show experimental results on three real world data sets, present related work, and close with a discussion of future ....

Tom M. Mitchell. Conditions for the equivalence of hierarchical and flat Bayesian classifiers. http://www.cs.cmu.edu/ tom/hierproof.ps, 1998.

.... often perform best [Joachims 1997; Nigam et al. 1998; McCallum Nigam 1998] Somewhat surprisingly, it can be shown that a probabilistic form of Pachinko Machine, when trained using maximum likelihood estimates and a constant vocabulary, is equivalent to the simple non hierarchical classifier [Mitchell 1998]. At each node in the hierarchy this non deterministic version of the Pachinko Machine assigns each document probabilistically to all of its descendants, whereas the deterministic Pachinko Machine proposed by Koller and Sahami assigns each document to its single most probable descendant. The ....

Mitchell, T. M. 1998. Conditions for the equivalence of hierarchical and flat Bayesian classifiers.

No context found.

T. Mitchell. Conditions for the equivalence of hierarchical and flat bayesian classifiers. Technical report, Center for Automated Learning and Discovery, Carnegie-Mellon University, 1998.

No context found.

T. Mitchell. Conditions for the equivalence of hierarchical and flat bayesian classifiers. Technical report, Center for Automated Learning and Discovery, Carnegie-Mellon University, 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