Ajou University repository

Gaussian Soft Decision Trees for Interpretable Feature-Based Classification
Citations

SCOPUS

0

Citation Export

Publication Year
2021-01-01
Journal
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Publisher
Springer Science and Business Media Deutschland GmbH
Citation
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol.12713 LNAI, pp.143-155
Keyword
Feature-based classificationGaussian mixturesGaussian Soft Decision TreesInterpretable machine learningTabular data
Mesh Keyword
Branched structuresDecision processFeature-basedFeature-based classificationGaining insightsGaussian mixturesInterpretabilityTree-structured
All Science Classification Codes (ASJC)
Theoretical Computer ScienceComputer Science (all)
Abstract
How can we accurately classify feature-based data such that the learned model and results are more interpretable? Interpretability is beneficial in various perspectives, such as in checking for compliance with exiting knowledge and gaining insights from decision processes. To gain in both accuracy and interpretability, we propose a novel tree-structured classifier called Gaussian Soft Decision Trees (GSDT). GSDT is characterized by multi-branched structures, Gaussian mixture-based decisions, and a hinge loss with path regularization. The three key features make it learn short trees where the weight vector of each node is a prototype for data that mapped to the node. We show that GSDT results in the best average accuracy compared to eight baselines. We also perform an ablation study of the various structures of covariance matrix in the Gaussian mixture nodes in GSDT and demonstrate the interpretability of GSDT in a case study of classification in a breast cancer dataset.
Language
eng
URI
https://aurora.ajou.ac.kr/handle/2018.oak/36661
https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85111090253&origin=inward
DOI
https://doi.org/10.1007/978-3-030-75765-6_12
Journal URL
https://www.springer.com/series/558
Type
Conference
Funding
Acknowledgments. Publication of this article has been funded by the Basic Science Research Program through the National Research Foundation of Korea (2018R1A1A3A0407953, 2018R1A5A1060031).
Show full item record

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Related Researcher

Lee, Sael Image
Lee, Sael이슬
Department of Software and Computer Engineering
Read More

Total Views & Downloads

File Download

  • There are no files associated with this item.