Generates a high-dimensional uniform domain and identifies points near the decision boundaries of multiple models.
Arguments
- models
A named list of fitted
classboundmodel objects.- data
A data frame representing the original training data. The bounding box of this data determines the simulation domain.
- n_points
Integer. Number of points to simulate in the high-dimensional space (default: 10000).
- threshold
Numeric. A percentile threshold (0 to 1) for determining boundary proximity. Points below this advantage percentile are considered boundary-near. Default: 0.05.
Value
A data frame containing only the simulated points that are near the boundary for at least one of the provided models. It has the simulated coordinates, the predicted classes for each model (prefixed with Y_), and a logical is_boundary column (always TRUE in the returned rows).
Details
This function follows the approach of the 'classifly' package (Wickham and Cook)
and Wickham, Cook and Hofmann (2015). Points are simulated uniformly within the
bounding box of the training data and classified by each model. For each point,
the "advantage" is computed as the difference between the two largest class
probabilities. Points whose advantage falls at or below the threshold
quantile are treated as lying near a decision boundary. When a model does not
return class probabilities, the advantage is approximated by the proportion of
the 5 nearest simulated neighbors (on range-scaled numeric features) that share
the point's predicted class.
References
Wickham, H., Cook, D. and Hofmann, H. (2015). Visualizing statistical models: Removing the blindfold. Statistical Analysis and Data Mining: The ASA Data Science Journal, 8(4), 203–225. doi:10.1002/sam.11271
Wickham, H. and Cook, D. classifly: Explore Classification Models in High Dimensions. R package. https://CRAN.R-project.org/package=classifly