Meta-Feature Taxonomy for Supporting Automatic Machine Learning

Many automatic machine learning (AutoML) libraries have been developed recently, meeting public demand for more machine learning tools which can be used without an expert. A common tactic illicited by these frameworks is to initially generate meta-features which are then used as an initial heuristic for further evaluation in recent AutoML frameworks. In this thesis we provide a systematic categorization of meta-features in the AutoML literature. Current implementations of automatic machine learning frameworks fail to provide reasoning for meta-feature selection, and a taxonomic categorization is needed. Our approach reviewed current AutoML frameworks and created a taxonomy of five categories into which any meta-feature can be categorized. We have created a general framework with which any currently used meta-features can be described, as well as demonstrate some scenarios for their applications. Additionally, a runtime analysis of the wall-clock time required for meta-feature generation is provided for 18 data collections found in previous CHALearn AutoML competitions, which took between 0:10:26.9, and 98:43:46.5. Additionally we found that a sample percentage of 0.1 is sufficient for use in Sample Variant Landmark Meta-Feature generation when using the Nearest Neighbour, Elite Nearest Neighbour, Best Decision Node, and Random Decision Node Landmarks which indicates potential use as meta-features in AutoML.

Name thesis:degree_name

Master of Science (MSc)

Discipline thesis:degree_discipline

Computer Science

Grantor dc:publisher.institution

Science

Year dc:date.issued

2019

Author dc:creator

• Davies, Cooper

Advisors dc:contributor.advisor

• Maurer, Frank
• Denzinger, Jörg

Committee members dc:contributor.committeemember

• Jacob, Christian
• Alim, Usman R.
• Oehlberg, Lora A.

• Machine Learning
• Meta-Feature
• Automatic Machine Learning
• AutoML

Statement dc:rights

• University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.

Language dc:language.iso

eng

DOI

http://dx.doi.org/10.11575/PRISM/37377

OAI identifier oai:identifier

oai:ucalgary.scholaris.ca:1880/111399