A Study of the Correlation of Metafeatures Used for Metalearning

Author(s)

Rivolli, Adriano ; Garcia, Luis P. F. ; Lorena, Ana C. ; de Carvalho, Andre C. P. L. F. ; Rojas, I ; Joya, G ; Catala, A

• Document type: Journal article
• Source: ADVANCES IN COMPUTATIONAL INTELLIGENCE, IWANN 2021, PT I; v. 12861, p. 13-pg., 2021-01-01.

Abstract

Metalearning has been largely used over the last years to recommend machine learning algorithms for new problems based on past experience. For such, the first step is the creation of metabase, or metadataset, containing metafeatures extracted from several datasets along with the performance of a pool of candidate algorithm(s). The next step is the induction of machine learning metamodels using the metabase as input. These models can recommend the most suitable algorithms for new datasets based on their metafeatures values. An effective metalearning system must employ metafeatures that characterize essential aspects of the datasets while also distinguishing different problems and solutions. The characterization process should also show a low computational cost, otherwise, the recommendation system can be replaced by a standard trial-and-error approach. This paper proposes the use of an unsupervised correlation-based feature selection strategy to identify a reduced subset of metafeatures for metalearning systems. Empirically, the predictive performance achieved by metalearning systems using the subset of selected metafeatures is similar or better than the performance obtained using the whole set of metafeatures. In addition, a noteworthy reduction in the number of metafeatures needed is observed, implying computational cost reductions. (AU)

• FAPESP's process: 13/07375-0 - CeMEAI - Center for Mathematical Sciences Applied to Industry
• Grantee: Francisco Louzada Neto
• Support Opportunities: Research Grants - Research, Innovation and Dissemination Centers - RIDC