Pfahringer B.: Controlling Constructive Induction in CiPF: An MDL Approach, in Proceedings of the European Conference on Machine Learning (ECML94), 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....unclassified examples. This phenomenon can be called overfitting the representation language in analogy to fitting the noise. The presence of both noise and an inadequate language obviously increases the possibilities for overfitting even further. A preliminary version of CiPF was described in [Pfahringer 94a] The system CiPF 2.0 described in this chapter has been considerably improved. These modifications allow for robust constructive induction handling both noise and an inadequate language. This chapter is based partly on [Pfahringer 94b] Section 6.2 briefly describes a generic architecture for ....

Pfahringer B.: Controlling Constructive Induction in CiPF: An MDL Approach, in Proceedings of the European Conference on Machine Learning (ECML94), 1994.

....to a large number of learning problems in the past. Examples include hand printed character recognition [14] decision tree induction [20] molecular evolution [1, 18] analysing dynamic systems [8] learning engineering models [21] clustering [6] computer vision [13] and constructive induction [19]. In the context of automated concept acquisition from linguistic corpora, however, applications of the MDL principle have been relatively few. MDL principle has been explored for lexical knowledge acquisition [2] speech segmentation [3, 7] and to phonology [10, 11] Conceptually, induction can ....

B. Pfahringer. Controlling constructive induction in ciPF: An MDL approach. In F. Bergadano and L. de Raedt, editors, Proceedings of the European Conference on Machine Learning, pages 242--256, Berlin, 1994. Springer.

....of the newly defined features or predicates in the complexity of the hypothesis we have to pay a price for defining a large number of new features or predicates. Otherwise we would overfit the potential noise in the data by means of the language, a problem termed language fitting in [Pfahringer 94] Pfahringer 94] also suggests the usefulness of a function measuring both accuracy and complexity of a hypothesis and the definitions of the new features or predicates. This can be done by a measure based on the Minimum Description Length (MDL) principle [Rissanen 78] that measures both ....

....defined features or predicates in the complexity of the hypothesis we have to pay a price for defining a large number of new features or predicates. Otherwise we would overfit the potential noise in the data by means of the language, a problem termed language fitting in [Pfahringer 94] Pfahringer 94] also suggests the usefulness of a function measuring both accuracy and complexity of a hypothesis and the definitions of the new features or predicates. This can be done by a measure based on the Minimum Description Length (MDL) principle [Rissanen 78] that measures both accuracy and complexity ....

[Article contains additional citation context not shown here]

Pfahringer B.: Controlling Constructive Induction in CiPF: An MDL Approach, in Bergadano F. & Raedt L.de(eds.), Machine Learning: ECML-94, Springer, Berlin, pp.242-256, 1994.

....poorly at predicting concept membership of unclassified examples. This phenomenon can be called overfitting the representation language in analogy to fitting the noise. The presence of both noise and an inadequate language obviously increases the possibilities for overfitting even further. In [Pfahringer 94] we reported on our constructive learner CiPF and its relative success in noise free domains due to rigid control applying the Minimum Description Length principle. This paper will concentrate on how the improvements found in the newest version CiPF 2.0 allow for robust constructive induction ....

Pfahringer B.: Controlling Constructive Induction in CiPF: An MDL Approach, in Proceedings of the European Conference on Machine Learning (ECML94), 1994.

....the original representation space into a space where the learning examples exhibit (more) regularities. Usually this is done by introducingnew attributes and forgetting old ones. So constructive induction is searching for an adequate representation language for the learning task at hand. In [Pfahringer 94a] we reported on our constructive learner CIPF and its relative success in noise free domains due to rigid control based on the Minimum Description Length [Rissanen 78] principle. The main improvements in CIPF 2.0 are the use of the well knownsophisticated decision tree learner C4.5 [Quinlan 93] ....

Pfahringer B.: Controlling Constructive Induction in CiPF: An MDL Approach, in Proceedings of the European Conference on Machine Learning (ECML94), 1994.

.... hypothesis language [Stahl, 1996] In particular, socalled wrapper approaches to constructive induction, where the theory learned in one iteration is analyzed for the construction of new features for subsequent iterations, might easily be cast into this framework [Wnek and Michalski, 1994; Pfahringer, 1994; Kramer, 1994] With some elaboration, a general algorithm akin to the one described above could also incorporate other procedures for dimensionality reduction, like wrapper approaches to feature subset selection or the improved windowing algorithms we have described in the previous section. In ....

Bernhard Pfahringer. Controlling constructive induction in CiPF: an MDL approach. In Pavel B. Brazdil, editor, Proceedings of the 7th European Conference on Machine Learning (ECML-94), Lecture Notes in Artificial Intelligence, pages 242--256, Catania, Sicily, 1994. Springer-Verlag.

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