Dimensionality Reduction for the Algorithm Recommendation Problem

Given the increase in data generation, as many algorithms have become available in recent years, the algorithm recommendation problem has attracted increasing attention in Machine Learning. This problem has been addressed in the Machine Learning community as a learning task at the meta-level where the most suitable algorithm has to be recommended for a specific dataset. Since it is not trivial to define which characteristics are the most useful for a specific domain, several meta-features have been proposed and used, increasing the meta-data meta-feature dimension. This study investigates the influence of dimensionality reduction techniques on the quality of the algorithm recommendation process. Experiments were carried out with 15 algorithm recommendation problems from the Aslib library, 4 meta-learners, and 3 dimensionality reduction techniques. The experimental results showed that linear aggregation techniques, such as PCA and LDA, can be used in algorithm recommendation problems to reduce the number of meta-features and computational cost without losing predictive performance.