A Many-Objective Optimization Approach to Generate Synthetic Datasets based on Real-World Classification Problems
Steffano Xavier Pereira, Pericles B. C. de Miranda, Thiago Rodrigues de França, Carmelo J. A. Bastos-Filho, Tapas Si · 2022
The performance of classification algorithms is commonly assessed by employing annotated datasets. However, analyzing specific aspects of classifiers might require datasets with certain desired features, and synthetic datasets can be a practical option for classifier evaluation since generation procedures can be parameterized to produce datasets with convenient features for each case. Generating synthetic datasets is usually modeled as an optimization problem where data complexity measures are adopted as objective functions to estimate the complexity of the dataset. This paper proposes using a many-objective algorithm to generate synthetic datasets from four different characteristics of real-world problems. The results showed that the proposal could optimize conflicting objectives, generating synthetic datasets of specific complexities from real-world classification problems.