1/20/2022: Automatic Binary Classification, Alexis Espinoza, Diego Henríquez and Catalina Pezo, University of Concepcion

Presentation description:

In this presentation, the speakers showed their preliminary work on Automatic ML Algorithm Selection for Binary Classification. For this, they showed how they collected and generated a number of binary classification (92 public and 4000 self-generated) instances (datasets) on which the task was to predict the correct label for the unlabeled entries, based on the labeled entries whose percentage, in each dataset, could vary. They ran 9 different classification algorithms with default parameters and trained a deep learning model as in [1], achieving a 54% accuracy on the testing dataset. Several follow-up ideas about how to generate adequate training data and characterize the instances were discussed. 

References: 

• [1] Loreggia, Andrea, et al. “Deep learning for algorithm portfolios.” Thirtieth AAAI Conference on Artificial Intelligence. 2016.
• [2] Jesus, Alexandre D., et al. “Algorithm selection of anytime algorithms.” Proceedings of the 2020 genetic and evolutionary computation conference. 2020.
• [3] Huerta, I. I., Neira, D. A., Ortega, D. A., Varas, V., Godoy, J., & Asín-Achá, R. (2021). Improving the state-of-the-art in the Traveling Salesman Problem: An Anytime Automatic Algorithm Selection. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2021.115948 (Published online)
• [4] P. Baumann, D. S. Hochbaum, Y. T. Yang. “A comparative study of the leading machine learning techniques and two new optimization algorithms”. European Journal of Operational Research, Volume 272, Issue 3, 1 February 2019, Pages 1041-1057