bayesRegress {mmb}

Perform full-dependency Bayesian regression for a sample.

Description

This method performs full-dependency regression by discretizing the continuous target variable into ranges (buckets), then finding the most probable ranges. It can either regress on the values in the most likely range or sample from all ranges, according to their likelihood.

Usage

bayesRegress(
  df,
  features,
  targetCol,
  selectedFeatureNames = c(),
  shiftAmount = 0.1,
  retainMinValues = 2,
  doEcdf = FALSE,
  useParallel = NULL,
  numBuckets = ceiling(log2(nrow(df))),
  sampleFromAllBuckets = TRUE,
  regressor = NULL
)

Arguments

df

data.frame that contains all the feature's data

features

data.frame with bayes-features. One of the features needs to be the label-column.

targetCol

string with the name of the feature that represents the label.

selectedFeatureNames

vector default c(). Vector of strings that are the names of the features the to-predict label depends on. If an empty vector is given, then all of the features are used (except for the label). The order then depends on the features' order.

shiftAmount

numeric an offset value used to increase any one probability (factor) in the full built equation. In scenarios with many dependencies, it is more likely that a single conditional probability becomes zero, which would result in the entire probability being zero. Since this is often useless, the 'shiftAmount' can be added to each factor, resulting in a non-zero probability that can at least be used to order samples by likelihood. Note that, with a positive 'shiftAmount', the result of this function cannot be said to be a probability any longer, but rather results in a comparable likelihood (a 'probability score').

retainMinValues

integer to require a minimum amount of data points when segmenting the data feature by feature.

doEcdf

default FALSE a boolean to indicate whether to use the empirical CDF to return a probability when inferencing a continuous feature. If false, uses the empirical PDF to return the rel. likelihood. This parameter does not have any effect if all of the variables are discrete or when doing a regression. Otherwise, for each continuous variable, the probability to find a value less then or equal - given the conditions - is returned. Note that the interpretation of probability using the ECDF much deviates and must be used with care, especially since it affects each factor in Bayes equation that is continuous. This is especially true for the case where shiftAmount > 0.

useParallel

default NULL a boolean to indicate whether to use a previously registered parallel backend. If no explicit value was given, calls foreach::getDoParRegistered() to check for a parallel backend. When using parallelism, this function calculates each factor in the numerator and denominator of the final equation in parallel.

numBuckets

integer the amount of buckets to for discretization. Buckets are built in an equidistant manner, not as quantiles (i.e., one bucket has likely a different amount of values than another).

sampleFromAllBuckets

default TRUE boolean to indicate how to obtain values for regression from the buckets. If true, than takes values from those buckets with a non-zero probability, and according to their probability. If false, selects all values from the bucket with the highest probability.

regressor

Function that is given the collected values for regression and thus finally used to select a most likely value. Defaults to the built-in estimator for the empirical PDF and returns its argmax. However, any other function can be used, too, such as min, max, median, average etc. You may also use this function to obtain the raw values for further processing.

Author(s)

Sebastian Hönel [e-mail]

Examples

w <- mmb::getWarnings()
mmb::setWarnings(FALSE)

df <- iris[, ]
set.seed(84735)
rn <- base::sample(rownames(df), 150)
dfTrain <- df[1:120, ]
dfValid <- df[121:150, ]
tf <- mmb::sampleToBayesFeatures(dfValid[1,], "Sepal.Length")
mmb::bayesRegress(dfTrain, tf, "Sepal.Length")

mmb::setWarnings(w)