Meta Learning Recommendation System for Classification
Abstract
A data driven approach is an emerging paradigm for the handling of analytic problems. In this paradigm the mantra is to let the data speak freely. However, when using machine learning algorithms, the data does not naturally reveal the best or even a good approach for algorithm choice. One method to let the algorithm reveal itself is through the use of Meta Learning, which uses the features of a dataset to determine a useful model to represent the entire dataset. This research proposes an improvement on the meta-model recommendation system by adding classification problems to the candidate problem space with appropriate evaluation metrics for these additional problems. This research predicts the relative performance of six machine learning algorithms using support vector regression with a radial basis function as the meta learner. Six sets of data of various complexity are explored using this recommendation system and at its best, the system recommends the best algorithm 67 of the time and a good algorithm from 67 to 100 of the time depending on how good is defined.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 26, 2020
- Accession Number
- AD1103676
Entities
People
- Clarence Iii O. Williams
Organizations
- Air Force Institute of Technology