sortinghat package is a framework in R to streamline the evaluation of
classifiers (classification models and algorithms) and seeks to determine the
best classifiers on a variety of simulated and benchmark data sets with a
collection of benchmark metrics.
You can install the stable version on CRAN:
install.packages('sortinghat', dependencies = TRUE)
If you prefer to download the latest version, instead type:
A primary goal of
sortinghat is to enable rapid benchmarking across a variety of
classification scenarios. To achieve this, we provide a large selection of both
real and simulated data sets collected from the literature and around the
sortinghat, researchers can quickly replicate findings within the
literature as well as rapidly prototype new classifiers.
The list of real and simulated data sets will continue to grow. Contributions are greatly appreciated as pull requests.
Benchmark data sets are useful for evaluating and comparing classifiers...
(Work in Progress: Version 0.2 will include a collection of benchmark data sets)
Simulated Data Sets
In addition to benchmark data sets,
sortinghat provide a large collection of
data-generating models for simulations based on studies in the literature. Thus
far, we have added multivariate simulation models based on the following family
- Multivariate Normal
- Multivariate Student's t
- Multivariate Contaminated Normal
- Multivariate Uniform
Moreover, data can be generated based on the well-known configurations from:
The simulated data sets listed above can be generated via the
Classifier superiority is often determined by classification error rate (1 - accuracy). To assess classification efficacy, we utilize the following error-rate estimators:
- Cross-validation Error Rate
- Bootstrap Error Rate
- .632 Estimator from Efron (1983)
- .632+ Estimator from Efron and Tibshirani (1997)
- Bootstrap Cross-validation from Fu, Carrol, and Wang (2005)
- Leave-One-Out Bootstrap Error Rate
- Apparent Error Rate
Each of these error rates can be accessed via the
errorest function, which
acts as a wrapper around the error-rate estimators listed above.