2021-09-21 16:52:53 - Atualizado em 2021-09-21 16:54:05

Ciência de Dados para Eleições: A Graph-Based Spatial Cross-Validation Approach to Understand Election Results

Por Ricardo Marcacini

This paper addresses the dependence issues inherent from spatial data by proposing a novel graph-based Spatial Cross Validation approach designed to assess models learned with selected features from spatially contextualized electoral datasets. Our procedure takes advantage of the spatial graph structure from lattice-type spatial objects to define each fold as a local training set by removing highly correlated and distant data that may interfere with error estimates. Experiments involving the second round of the 2018 Brazilian presidential election demonstrate that our approach defines buffer regions that decrease the spatial dependence between training and test samples and filters the training set based on the spatial closeness of the test fold to provide more local and realistic data modeling.

Autores: Tiago Pinho da Silva, Antonio R. S. Parmezan, Gustavo E. A. P. A. Batista

Venue: ICMLA 2021 (IEEE INTERNATIONAL CONFERENCE ON MACHINE LEARNING AND APPLICATIONS)