Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

Understanding time series networks: a case study in rule extraction

M W Craven1, J W Shavlik

  • 1Computer Sciences Department, University of Wisconsin-Madison, 53706, USA. mark.craven@cs.cmu.edu

International Journal of Neural Systems
|August 1, 1997
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Biomedical informatics training at the University of Wisconsin-Madison.

Yearbook of medical informatics·2007
Same author

Evaluating machine learning approaches for aiding probe selection for gene-expression arrays.

Bioinformatics (Oxford, England)·2002
Same author

Identification of toxicologically predictive gene sets using cDNA microarrays.

Molecular pharmacology·2001
Same author

Neural network input representations that produce accurate consensus sequences from DNA fragment assemblies.

Bioinformatics (Oxford, England)·1999
Same author

Increasing consensus accuracy in DNA fragment assemblies by incorporating fluorescent trace representations.

Proceedings. International Conference on Intelligent Systems for Molecular Biology·1997
Same author

Improving the quality of automatic DNA sequence assembly using fluorescent trace-data classifications.

Proceedings. International Conference on Intelligent Systems for Molecular Biology·1996

We developed TREPAN to create understandable decision trees from neural networks. TREPAN extracts accurate and interpretable models, outperforming traditional methods for tasks like exchange rate prediction.

Area of Science:

  • Artificial Intelligence
  • Machine Learning
  • Computational Finance

Background:

  • Neural networks often create complex, uninterpretable models.
  • Extracting symbolic representations from neural networks is a significant challenge.
  • Understanding model decisions is crucial for trust and validation, especially in finance.

Purpose of the Study:

  • To introduce TREPAN, an algorithm for generating comprehensible symbolic representations from trained neural networks.
  • To evaluate TREPAN's effectiveness in extracting accurate and interpretable decision trees.
  • To compare TREPAN's extracted trees against conventionally induced decision trees.

Main Methods:

  • Developed the TREPAN algorithm to extract decision trees from trained neural networks.

Related Experiment Videos

  • Applied TREPAN to a neural network trained for predicting the Dollar-Mark exchange rate.
  • Conducted comparative experiments using conventional decision tree induction algorithms on the same training data.

    • Main Results:

    • TREPAN successfully extracted a decision tree that matched the neural network's predictive accuracy.
    • The extracted decision tree provided a comprehensible representation of the learned concept.
    • Decision trees induced directly from training data showed lower accuracy and comprehensibility compared to TREPAN's output.

    Conclusions:

    • TREPAN offers a viable method for creating interpretable models from complex neural networks.
    • The algorithm enhances the practical utility of neural networks by making their learned representations understandable.
    • TREPAN-extracted models are superior to conventionally induced trees for noisy time-series prediction tasks.