N. Japkowicz. Concept-Learning in the absence of counterexamples: an autoassociation-based approach to classification. PhD thesis, New Brunswick Rutgers, The State University of New Jersey, 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....useful when the data is distributed in subspaces. 1. Introduction The goal of the Data Description is to distinguish between a target set of objects and all other possible objects. It is also called One Class classification[6] Outlier Detection[7] Novelty detection[1] or Concept learning[4]. It is mainly used to detect new objects that resemble a known set of objects. When the object does not resemble the data, it is likely an outlier or a novelty. When it is accepted by the data description, it can be classified with more confidence in a subsequent classifier. Different methods ....

N. Japkowicz. Concept-Learning in the absence of counterexamples: an autoassociation-based approach to classification. PhD thesis, New Brunswick Rutgers, The State University of New Jersey, 1999.

....Gaussians and the Parzen density estimation. The second type of methods fit a model to the data and compute the distance # T (x) to this model. Here we will use four simple models, the support vector data description [14] k means clustering, k center method [15] and an autoencoder neural network [6]. Here a descriptive model is fitted to the data and the resemblance (or distance) to this model is used. In the SVDD a hypersphere is put around the data. By applying the kernel trick (analogous to the support vector classifier) the model becomes more flexible to follow the characteristics in the ....

N. Japkowicz. Concept-Learning in the absence of counter-examples: an autoassociation-based approach to classification. PhD thesis, New Brunswick Rutgers, The State University of New Jersey, 1999.

.... type illustrated by the domain of Figure 6(e) 11 This study suggested that no relationship between an increase in problem size (with a constant class imbalance ratio) and an increase in classification error could be established for either the autoassociator or the MLP network in such problems (Japkowicz, 2000b) 6. Suggested Framework for Fine Tuning the Current Results The observations made in the previous section will now be used to explain the results obtained on real world domains in Section 3.3 and suggest new research directions to fine tune our current results in case of a mismatch. This ....

Japkowicz, N. (1999a). Concept-Learning in the Absence of Counter-Examples: An AutoassociationBased Approach to Classification, Ph.D. Dissertation, Technical Report DCS-TR-390, Rutgers University.

....for unimodal data, then every point in the training set activates hidden unit values located in the same vicinity and which increase or decrease monotonically. Such data 12 We would like to thank Gary Cottrell for suggesting this comparison. 13 Support for this speculation can be found in [Japkowicz1999] which shows a clear decrease in classification accuracy by NL NL as the conceptual clusters of the domain in Figure 1(c) are moved closer to one another. can, thus, be approximated linearly, in the same fashion as they would be by a linear autoassociator. On the other hand, if the data set is ....

Nathalie Japkowicz. Concept-Learning in the Absence of Counter-Examples: An Autoassociation-Based Approach to Classification. PhD thesis, Rutgers University, October 1999.

....the core of this paper. It begins by discussing the experiments designed to analyze and compare the operation of the two systems, it then presents the results of this analysis, and it concludes with a discussion of our results. Section 4 summarizes our findings. Additional details can be found in [3]. 1 Another interesting aspect of this dichotomy is that the recognition based system can be significantly more accurate than the discrimination based one (see [4] 3] 2] 2 Discrimination versus Recognition We begin this section by describing the two classification systems contrasted in ....

....and it concludes with a discussion of our results. Section 4 summarizes our findings. Additional details can be found in [3] 1 Another interesting aspect of this dichotomy is that the recognition based system can be significantly more accurate than the discrimination based one (see [4] [3], 2] 2 Discrimination versus Recognition We begin this section by describing the two classification systems contrasted in this study and we then present the efficiency results they obtained on an abstract problem which displays nonlinearly separable and multi modal properties. 2.1 Systems ....

[Article contains additional citation context not shown here]

Nathalie Japkowicz. Concept-Learning in the Absence of Counter-Examples: An Autoassociation-Based Approach to Classification (in progress). PhD thesis, Rutgers University, 1999.

....which negative data (gearbox or motor failure) are scarce. 9 Its second feature is that the absence of negative data during the learning phase yields a bias different from those created by classifiers which use both classes of data for training. This question is currently being investigated in [Japkowicz1999] Acknowledgments We would like to express our gratitude to Gary Cottrell and an anonymous reviewer for their extremely helpful comments. We would also like to thank Inna Stainvas for useful discussions about autoassociators and for her careful reading of this manuscript. This research was ....

Nathalie Japkowicz. Concept-Learning in the Absence of Counter-Examples: An Autoassociation-Based Approach to Classification (in progress). PhD thesis, Rutgers University, 1999.

....concepts while specializing using their counterexamples. Although recognition based learners are also able to generalize from examples of the concept, they do not, however, have access to counterexamples for specialization. Their ability to specialize thus comes from certain internal capabilities ([Japkowicz et al..1999] discusses these capabilities in detail for the autoassociator) The aforementioned contrast between discrimination and recognition based classification methods suggests that differences in the classification performance of the two types of systems may be introduced by domains requiring ....

....:9) 5; 2) 2; 5) 5; 8) and ( 8; 5) As shown in Figure 2(a) each mean actually stands for a cluster of data points normally distributed around that mean and of variance oe 2 = 0:01. This domain was inspired from the real world domains used in [Japkowicz et al..1995] as described in [Japkowicz1999] Feature 1 Feature # 2 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Feature # Feature # 2 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Number of Negative Misclass Percent of Positive Accepted 100 200 300 400 20 40 60 80 100 (a) Neutral (b) Means Only (c) ROC Analysis Figure 2: a) ....

[Article contains additional citation context not shown here]

Nathalie Japkowicz. ConceptLearning in the Absence of Counter-Examples: An Autoassociation-Based Approach to Classification (in progress). PhD thesis, Rutgers University, 1999.

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