This vignette demonstrates how to use the auditor package for auditing residuals of models. The auditor provides methods for model verification and validation by error analysis.

Many models, such as random forests and neutral networks are nowadays treated as black boxes. Therefore, there is a lack of theory that describes the behavior of errors in those models. Most methods provided in auditor package are model-agnostic, so can be used regardless of knowledge about errors.

Some of the graphical error analysis methods also have corresponding scores, which allow comparison of two models.

Regression use case - dragons data

To illustrate applications of auditor to regression problems we will use an artificial dataset dragons available in the DALEX package. Our goal is to predict the length of life of dragons.

library(DALEX)
data("dragons")
head(dragons)
##   year_of_birth   height   weight scars colour year_of_discovery
## 1         -1291 59.40365 15.32391     7    red              1700
## 2          1589 46.21374 11.80819     5    red              1700
## 3          1528 49.17233 13.34482     6    red              1700
## 4          1645 48.29177 13.27427     5  green              1700
## 5            -8 49.99679 13.08757     1    red              1700
## 6           915 45.40876 11.48717     2    red              1700
##   number_of_lost_teeth life_length
## 1                   25   1368.4331
## 2                   28   1377.0474
## 3                   38   1603.9632
## 4                   33   1434.4222
## 5                   18    985.4905
## 6                   20    969.5682

Models

Linear model

lm_model <- lm(life_length ~ ., data = dragons)

Random forest

library("randomForest")
set.seed(59)
rf_model <- randomForest(life_length ~ ., data = dragons)

Preparation for error analysis

The beginning of each analysis is creation of an explainer object with DALEX package. It’s an object that can be used to audit a model.

lm_exp <- DALEX::explain(lm_model, label = "lm", data = dragons, y = dragons$life_length)
## Preparation of a new explainer is initiated
##   -> model label       :  lm 
##   -> data              :  2000  rows  8  cols 
##   -> target variable   :  2000  values 
##   -> predict function  :  yhat.lm  will be used (default)
##   -> predicted values  :  numerical, min =  540.9447 , mean =  1370.986 , max =  3925.691  
##   -> residual function :  difference between y and yhat (default)
##   -> residuals         :  numerical, min =  -108.2062 , mean =  -3.701928e-12 , max =  113.8603  
## A new explainer has been created!
rf_exp <- DALEX::explain(rf_model, label = "rf", data = dragons, y = dragons$life_length)
## Preparation of a new explainer is initiated
##   -> model label       :  rf 
##   -> data              :  2000  rows  8  cols 
##   -> target variable   :  2000  values 
##   -> predict function  :  yhat.randomForest  will be used (default)
##   -> predicted values  :  numerical, min =  610.9752 , mean =  1370.181 , max =  3292.296  
##   -> residual function :  difference between y and yhat (default)
##   -> residuals         :  numerical, min =  -135.4756 , mean =  0.8047108 , max =  720.0888  
## A new explainer has been created!

Audit of residuals

In this section we give short overview of a visual validation of model errors and show the propositions for the validation scores. Auditor helps to find answers for questions that may be crucial for further analyses.

In further sections we will overview auditor functions for analysis of model residuals. They are discussed in alphabetical order.

model_residual()

In this vignette we use first pipeline. However, alternative evaluations are showed as comments. First, we need to create a modelResiduals objects.

library(auditor)
lm_mr <- model_residual(lm_exp)
rf_mr <- model_residual(rf_exp)

Plots

plot_acf() - Autocorrelation Function of Residuals

Autocorrelation Function plot can be used to check randomness of errors. If random, autocorrelations should be near zero for lag separations. If non-random, then autocorrelations will be significantly non-zero.

Residuals may be ordered by values of any model variable or by fitted values. If variable is not specified, function takes order from the data set.

plot(lm_mr, type = "acf", variable = "year_of_discovery")

plot of chunk unnamed-chunk-6

# alternative:
# plot_acf(lm_mr, variable = "year_of_discovery")

plot_autocorrelation() - Autocorrelation of Residuals

On the Autocorrelation plot there are i-th vs i+1-th residuals. This plot may be useful for checking autocorrelation of residuals.

Sometimes it is difficult to compare two models basing only on visualizations. Therefore, we have proposed some scores, which may be useful for choosing a better model.

plot(rf_mr, type = "autocorrelation")

plot of chunk unnamed-chunk-7

# alternative:
# plot_autocorrelation(rf_mr)

DW score and Runs score are based on Durbin-Watson and Runs test statistics.

Scores can be calculated with the scoreDW() and scoreRuns() functions or the score() function with argument score equals to “DW” or “Runs”.

rf_score_dw <- score_dw(rf_exp)
rf_score_runs <- score_runs(rf_exp)

rf_score_dw$score
## [1] 1.951918
rf_score_runs$score
## [1] -1.881788

plot_correlation - Correlation of Models

A grid of plots presents correlation of dependennt variable and fitted model values.

plot(rf_mr, lm_mr, type = "correlation")

plot of chunk unnamed-chunk-9

# alternative:
# plotM_correlation(rf_mr, lm_mr)

plot_pca() - Model PCA

Principal Component Analysis of models residuals. PCA can be used to assess the similarity of the models.

plot(rf_mr, lm_mr, type = "pca")

plot of chunk unnamed-chunk-10

# alternative:
# plot_pca(rf_audit, lm_audit)

plot_predition() - Observed vs Predicted

Basic plot of residuals vs observed, fitted or variable values. If variable is not specified, function takes order from the data set.

Black line corresponds to the y=x function.

plot(rf_mr, lm_mr, variable = "life_length", type = "prediction")

plot of chunk unnamed-chunk-11

# alternative:
# plot_prediction(rf_audit, lm_audit, variable = "life_length")

Predictions may be ordered by values any model variable or by fitted values. And both models may be plotted together.

plot(rf_mr, lm_mr, type = "prediction")

plot of chunk unnamed-chunk-12

# alternative:
# plot_prediction(rf_mr, lm_mr)

plot_rec() - Regression Error Characteristic (REC) Curve

Error Characteristic curves are a generalization of ROC curves. On the x axis of the plot there is an error tolerance and on the y axis there is a percentage of observations predicted within the given tolerance. REC curve estimates the Cumulative Distribution Function (CDF) of the error. Area Over the REC Curve (REC) is a biased estimate of the expected error.

plot(rf_mr, lm_mr, type = "rec")

plot of chunk unnamed-chunk-13

# alternative:
# plot_rec(rf_mr, lm_mr)

plot_residual() - Plot Residuals vs Observed, Fitted or Variable Values

Basic plot of residuals vs observed, fitted or variable values. It provides information about the structure of the model.

plot(rf_mr, type = "residual")

plot of chunk unnamed-chunk-14

# alternative:
# plot_residual(rf_mr)

Residuals may be ordered by values any model variable or by fitted values. And both models may be plotted together. If variable is not specified, function takes order from the data set.

plot(rf_mr, lm_mr, type = "residual", variable = "_y_hat_")

plot of chunk unnamed-chunk-15

# alternative:
# plot_residual(rf_mr, lm_mr, variable = "_y_hat_")

plotResidualBoxplot() - Boxplot of residuals

Comparison of the absolute valued of residuals. The red dot stands for the root mean square.

plot(lm_mr, rf_mr, type = "residual_boxplot")

plot of chunk unnamed-chunk-16

# alternative
# plot_residual_boxplot(lm_mr, rf_mr)

plot_residual_density() - Density of Residuals

Density of residuals may be plotted in different ways. Residuals of models may be simply compared.

plot(rf_mr, lm_mr, type = "residual_density")

plot of chunk unnamed-chunk-17

# alternative
# plot_residual_density(rf_mr, lm_mr)

Resuduals may be also divided by median of the numeric variable and splitted by a factor variable

plot_residual_density(rf_mr, lm_mr, variable = "life_length")

plot of chunk unnamed-chunk-18

plot_rroc() - Regression Receiver Operating Characteristic (RROC)

The basic idea of the ROC curves for regression is to show model asymmetry. The RROC is a plot where on the x-axis we depict total over-estimation and on the y-axis total under-estimation.

For RROC curves we use a shift, which is an equvalent to the threshold for ROC curves. For each observation we calculate new prediction: \eqn{\hat{y}'=\hat{y}+s} where s is the shift. Therefore, there are different error values for each shift: \eqn{e_i = \hat{y_i}' - y_i}

Over-estimation is caluclates as: \eqn{OVER= \sum(e_i|e_i>0)}. Under-estimation is calculated as: \eqn{UNDER = \sum(e_i|e_i<0)}. The shift equals 0 is represented by a dot.

The Area Over the RROC Curve (AOC) equals to the variance of the errors multiplied by \eqn{frac{n2}{2}}.

plot(rf_mr, lm_mr, type = "rroc")

plot of chunk unnamed-chunk-19

# alternative:
# plot_rroc(rf_mr, lm_mr)

plot_scalelocation() - Scale Location Plot

This plot shows if residuals are spread equally along the ranges of predictors.

plot(rf_mr, lm_mr, type = "scalelocation")

plot of chunk unnamed-chunk-20

# alternative:
# plot_scalelocation(rf_mr, lm_mr)

For comparing 2 models we can use GQ score, which is based on Goldfeld-Quandt test statistic. And may be computed also in score() function with argument score equals “GQ”.

plot_tsecdf() - Two-sided Empirical Cumulative Distribution Function (ECDF)

Cumulative Distribution Function for positive and negative residuals.

plot(rf_mr, lm_mr, type = "tsecdf")

plot of chunk unnamed-chunk-21

# alternative
# plot_tsecdf(rf_audit, lm_audit)

Other methods

Other methods and plots are described in vignettes: