Robustly Fit Naive Bayes Classifier

Computes the conditional a-posterior probabilities of a categorical class variable given independent predictor variables using the Bayes rule. Parameter estimates are robustly calculated using approximations of the error function for a Gaussian density, see helpr::fit_gauss().

Usage

fit_nb(x, ...)

# Default S3 method
fit_nb(x, y, mad = FALSE, laplace = 0, keep.data = FALSE, ...)

# S3 method for class 'formula'
fit_nb(formula, data, ...)

# S3 method for class 'tr_data'
fit_nb(x, ...)

# S3 method for class 'libml_nb'
print(x, ...)

# S3 method for class 'libml_nb'
predict(
  object,
  newdata,
  type = c("raw", "class", "posterior"),
  threshold = NULL,
  ...
)

# S3 method for class 'libml_nb'
plot(
  x,
  data,
  features,
  plot_type = c("pdf", "cdf", "log_odds"),
  x_lab = "value",
  id,
  ...
)

# S3 method for class 'naiveBayes'
plot(
  x,
  data,
  features,
  plot_type = c("pdf", "cdf", "log_odds"),
  x_lab = "value",
  id,
  ...
)

Arguments

x

A numeric matrix, a tr_data class objects, or a data frame of predictors. If called from an S3 generic method (e.g. plot.libml_nb()) or print.libml_nb()), either a libml_nb or naiveBayes object.

...

Additional arguments passed to the default fit_nb() default method. Currently not used in the predict or print S3 methods, but is used in the S3 plot method, arguments passed to SomaPlotr::plotCDFlist() or SomaPlotr::plotPDFlist().

y

factor(n). If not passing a formula, a factor with true class names for each sample (row) in x.

mad

logical(1). Should non-parametric approximations be applied during the parameter estimation procedure. See Details.

laplace

double(1). Positive controlling Laplace smoothing. The default (0) disables Laplace smoothing.

keep.data

Logical. Should the training data used to fit the model be included in the model object? When building thousands of models, this can become a memory issue and thus the default is FALSE.

formula

A model formula of the form: class ~ x1 + x2 + ... (no interactions).

data

A data frame of predictors (categorical and/or numeric), i.e. the data frame used to train the model.

object

A libml_nb model object.

newdata

A data.frame with new predictors, containing at least the model covariates (possibly more columns than the training data). Note that the column names of newdata are matched against the training data.

type

character(1). Matched parameter. If "class", the class name with maximal posterior probability is returned for each sample, otherwise the conditional a-posterior probabilities for each class are returned. If called from the S3 plot method, a character type is a string determining the plot type, currently either CDF or PDF (default).

threshold

numeric(1). Indicating the minimum probability a prediction can take.

features

An optional feature specifying which subset of model features to plot. If missing, all features are plotted.

plot_type

character(1). A string determining the plot type, currently either a probability density function (PDF, default), CDF, or log-odds plots. Matched via match.arg().

x_lab

character(1). Optional label for the x-axis.

id

integer(n) or character(n). Optional identifier of a specific sample to plot on top of either PDFs or CDFs. Either an index of the sample row in the data, or its rowname.

Value

libml_nb: A naive Bayes model with robustly fit parameters.

predict.libml_nb: depending on the type argument, either the predicted class name or the posterior probability of a robustly estimated naive Bayes model.

plot.libml_nb: A plot, either a list of PDFs/CDFs, or a log-odds plot.

Details

When mad = TRUE (median absolute deviation), non-parametric calculation of Bayes' parameters are estimated, namely, mu = median(x) and sd = IQR(x) / 1.349. That is, fit_gauss(..., mad = TRUE).

Methods (by class)

• fit_nb(default): S3 default method for fit_nb.

• fit_nb(formula): S3 formula method for fit_nb.

• fit_nb(tr_data): S3 tr_data method for fit_nb.

Functions

• print(libml_nb): S3 print method for libml_nb.

• predict(libml_nb): S3 predict method for libml_nb.

• plot(libml_nb): S3 plot method for libml_nb.

• plot(naiveBayes): Plot a naiveBayes (e1071) model object.

References

This function was heavily influenced by e1071::naiveBayes() See David Meyer <email: David.Meyer@R-project.org>.

See also

helpr::fit_gauss()

plot_log_odds(), SomaPlotr::plotPDFlist(), SomaPlotr::plotCDFlist()

The fit*() family: fit_gbm(), fit_kknn(), fit_logistic()

Author

Stu Field

Examples

head(tr_iris)
#> ══ Training Data Object ═══════════════════════════════════════════════
#> • response        Species
#> • class labels    'setosa', 'virginica'
#> • counts          [50, 50]
#> • factor          TRUE
#> • n               2
#> ───────────────────────────────────────────────────────────────────────
#> # A tibble: 6 × 5
#>   Sepal.Length Sepal.Width Petal.Length Petal.Width Species  
#>          <dbl>       <dbl>        <dbl>       <dbl> <fct>    
#> 1         4.16        3.47         1.77       0.483 setosa   
#> 2         7.25        3.85         5.47       2.16  virginica
#> 3         6.82        3.20         6.35       1.96  virginica
#> 4         6.81        2.33         5.44       1.25  virginica
#> 5         7.31        2.58         4.77       3.26  setosa   
#> 6         5.72        2.83         2.26      -0.248 setosa   
# standard naive Bayes
m1 <- e1071::naiveBayes(Species ~ ., data = tr_iris) # non-robust params
m2 <- fit_nb(Species ~ ., data = tr_iris)  # formula syntax
m3 <- fit_nb(tr_iris)                      # tr_data syntax
m4 <- data.frame(tr_iris[, -5L]) |>
  fit_nb(y = tr_iris$Species)              # default syntax

# not same
identical(sapply(m1$tables, as.numeric), sapply(m2$tables, as.numeric))
#> [1] FALSE

# same
identical(sapply(m2$tables, as.numeric), sapply(m3$tables, as.numeric))
#> [1] TRUE

# Predictions
table(predict(m1, iris, type = "class"), iris$Species) # benchmark
#>            
#>             setosa versicolor virginica
#>   setosa        50          2         0
#>   virginica      0         48        50
table(predict(m2, iris, type = "class"), iris$Species) # approx same (Gaussian data)
#>            
#>             setosa versicolor virginica
#>   setosa        50          3         0
#>   virginica      0         47        50

# Plotting
plot(m2, tr_iris)
#> Warning: Non-positive values detected ...
#> If RFU data, possibly multi-logging?
#> [[1]]

#> 
#> [[2]]

#> 
#> [[3]]

#> 
#> [[4]]

#> 
plot(m2, tr_iris, id = 20)      # sample 20 is definitely "virginica"
#> Warning: Non-positive values detected ...
#> If RFU data, possibly multi-logging?
#> [[1]]

#> 
#> [[2]]

#> 
#> [[3]]

#> 
#> [[4]]

#> 
plot(m1, tr_iris, plot_type = "cdf")  # plot type CDF
#> Warning: Non-positive values detected ...
#> If RFU data, possibly multi-logging?
#> [[1]]

#> 
#> [[2]]

#> 
#> [[3]]

#> 
#> [[4]]

#> 
plot(m2, tr_iris, features = "Sepal.Length", id = 70)  # 1 feature
#> [[1]]

#> 
plot(m1, tr_iris, plot_type = "cdf", lty = "longdash") # pass-through of lty
#> Warning: Non-positive values detected ...
#> If RFU data, possibly multi-logging?
#> [[1]]

#> 
#> [[2]]

#> 
#> [[3]]

#> 
#> [[4]]

#>