Package: NumericEnsembles 0.5.0

NumericEnsembles: Automatically Runs 23 Individual and 17 Ensembles of Models

Automatically runs 23 individual models and 17 ensembles on numeric data. The package automatically returns complete results on all 40 models, 25 charts, multiple tables. The user simply provides the data, and answers a few questions (for example, how many times would you like to resample the data). From there the package randomly splits the data into train, test and validation sets, builds models on the training data, makes predictions on the test and validation sets, measures root mean squared error (RMSE), removes features above a user-set level of Variance Inflation Factor, and has several optional features including scaling all numeric data, four different ways to handle strings in the data. Perhaps the most significant feature is the package's ability to make predictions using the 40 pre trained models on totally new (untrained) data if the user selects that feature. This feature alone represents a very effective solution to the issue of reproducibility of models in data science. The package can also randomly resample the data as many times as the user sets, thus giving more accurate results than a single run. The graphs provide many results that are not typically found. For example, the package automatically calculates the Kolmogorov-Smirnov test for each of the 40 models and plots a bar chart of the results, a bias bar chart of each of the 40 models, as well as several plots for exploratory data analysis (automatic histograms of the numeric data, automatic histograms of the numeric data). The package also automatically creates a summary report that can be both sorted and searched for each of the 40 models, including RMSE, bias, train RMSE, test RMSE, validation RMSE, overfitting and duration. The best results on the holdout data typically beat the best results in data science competitions and published results for the same data set.

Authors:Russ Conte [aut, cre, cph]