Package {chaidr}

chaidr: CHAID and Exhaustive CHAID Decision Trees

Description

An implementation in base 'R' of the CHAID (Chi-squared Automatic Interaction Detection) decision tree algorithm of Kass (1980) doi:10.2307/2986296 and the Exhaustive CHAID variant of Biggs, de Ville, and Suen (1991) doi:10.1080/02664769100000005, as specified in the 'IBM SPSS' Statistics Algorithms documentation. Supports nominal, ordinal (with floating missing category), and continuous predictors, and nominal, ordinal, and continuous response variables using Pearson chi-squared, Goodman row-effects, and one-way ANOVA F tests respectively. Includes prediction, rule extraction, gains and lift analysis, validation on holdout data, and visualization via base graphics, 'Graphviz' DOT, 'plotly', and conversion to 'partykit' objects.

Author(s)

Maintainer: Osamu Morimoto galactic.supermarket@gmail.com [copyright holder]

Authors:

• Osamu Morimoto galactic.supermarket@gmail.com [copyright holder]

See Also

Useful links:

• https://github.com/morimotoosamu/chaidr

• https://morimotoosamu.github.io/chaidr/

• Report bugs at https://github.com/morimotoosamu/chaidr/issues

Fit a CHAID or Exhaustive CHAID decision tree

Description

Grows a CHAID (Chi-squared Automatic Interaction Detection) or Exhaustive CHAID decision tree following the IBM SPSS Statistics algorithm specification. Nominal (factor), ordinal (ordered) and continuous (numeric) predictors are supported; continuous predictors are discretised into quantile bins first. The response may be nominal (Pearson or likelihood-ratio chi-squared test), ordinal (Goodman row effects test) or continuous (one-way ANOVA F test).

Usage

chaid(
  formula,
  data,
  weights = NULL,
  freq = NULL,
  method = c("chaid", "exhaustive"),
  control = chaid_control(),
  costs = NULL,
  y_scores = NULL
)

Arguments

Details

Missing predictor values are handled as in SPSS: for ordinal predictors they form a floating category that may merge with any group, for nominal predictors they form an ordinary extra category. Cases with a missing response, missing/zero/negative weights, or all predictors missing are dropped before fitting.

Value

An object of class "chaid": a list with components call, method, control, response (name, type, levels, scores), predictors (internal coding of each predictor), nodes (list of node records with distribution, prediction and split information), costs, n (number of cases used) and n_dropped (number of excluded cases).

References

Kass, G. V. (1980). An exploratory technique for investigating large quantities of categorical data. Applied Statistics, 29(2), 119-127.

Biggs, D., de Ville, B., & Suen, E. (1991). A method of choosing multiway partitions for classification and decision trees. Journal of Applied Statistics, 18(1), 49-62.

See Also

chaid_control(), predict.chaid(), chaid_table(), chaid_rules(), plot.chaid()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
print(fit)
predict(fit, head(iris))

Convert a CHAID tree to a partykit constparty

Description

Converts a fitted "chaid" object to a partykit::constparty object so that the 'partykit' toolbox (plot(), print(), nodeapply(), 'ggparty', ...) can be used. Binned continuous predictors are represented as ordered factors of the bin interval labels, and all splits become index-type partykit::partysplit objects, so missing values ("<NA>" level) are routed explicitly and split rules display the interval labels.

Usage

chaid_as_party(x, data, weights = NULL, freq = NULL, ...)

## S3 method for class 'chaid'
as.party(obj, data, weights = NULL, freq = NULL, ...)

Arguments

Details

The returned object is intended for visualisation and structural inspection. For predictions on new data use predict.chaid(); the predict() method of the party object expects the converted data representation, not the original one.

Value

A partykit::constparty object.

See Also

chaid(), predict.chaid()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
pt <- chaid_as_party(fit, data = iris)
print(pt)

Control parameters for CHAID tree growing

Description

Collects the algorithm parameters used by chaid(). The defaults match the defaults of the IBM SPSS Statistics user interface (alpha 0.05, Bonferroni adjustment on, no re-splitting, maximum depth 3, minimum parent size 100, minimum child size 50, 10 bins for continuous predictors).

Usage

chaid_control(
  alpha_merge = 0.05,
  alpha_split = 0.05,
  alpha_split_merge = 0.05,
  resplit = FALSE,
  bonferroni = TRUE,
  stat = c("pearson", "lr"),
  exhaustive_adjust = c("spss", "biggs"),
  max_depth = 3L,
  min_parent = 100,
  min_child = 50,
  min_segment = NULL,
  n_bins = 10L,
  epsilon = 0.001,
  max_iter = 100L,
  adjust_across = c("none", "bonferroni", "holm", "hochberg", "hommel", "BH", "BY")
)

Arguments

Value

An object of class "chaid_control": a list of the validated parameter values, to be passed to the control argument of chaid().

See Also

chaid()

Examples

ctl <- chaid_control(max_depth = 2, min_parent = 20, min_child = 5)
fit <- chaid(Species ~ ., data = iris, control = ctl)
fit

Render a CHAID tree as Graphviz DOT

Description

chaid_dot() generates a publication-quality Graphviz DOT description of the tree using base R only. It can be rendered with chaid_graphviz() (which uses 'DiagrammeR', bundling viz.js so no Graphviz binary is needed) or written to a .gv file and rendered externally with dot -Tpng / dot -Tsvg.

Usage

chaid_dot(
  fit,
  palette = NULL,
  rankdir = "TB",
  label_len = 28,
  legend = TRUE,
  file = NULL
)

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

chaid_graphviz(fit, ...)

Arguments

Value

For chaid_dot(), the DOT source as a character string of class "chaid_dot" (returned invisibly; its print() method outputs the source). For chaid_graphviz(), an 'htmlwidget' as returned by DiagrammeR::grViz().

See Also

plot.chaid(), chaid_plotly()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
dot <- chaid_dot(fit)

Gains and lift table for a CHAID tree

Description

Sorts the terminal nodes by decreasing response rate (node mean for continuous responses) and computes cumulative gains and lift, corresponding to the SPSS gains table.

Usage

chaid_gains(fit, data = NULL, target = NULL, weights = NULL, freq = NULL)

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

## S3 method for class 'chaid_gains'
plot(x, type = c("gains", "lift"), ...)

Arguments

Value

An object of class "chaid_gains", a list with the gains table (table), the target level, the response type (ytype), the overall rate (overall) and the data source ("training" or "newdata"). print() and plot() methods are available.

See Also

chaid_table(), chaid_validate()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
g <- chaid_gains(fit, target = "virginica")
print(g)
plot(g)
plot(g, type = "lift")

Heuristic variable importance for a CHAID tree

Description

SPSS defines no official importance measure for CHAID, and the chi-squared and F statistics of different splits are not directly comparable because their degrees of freedom differ. This function therefore uses a p-value based heuristic: for each predictor, importance is the sum over its splits of ⁠(node size / root size) * (-log10(adjusted p-value))⁠, i.e. a variable is important when it splits large nodes with strong significance.

Usage

chaid_importance(fit)

Arguments

Value

A data frame with one row per predictor actually used for a split, sorted by decreasing importance: variable, n_splits, min_p_adj, importance and importance_pct (share of the total importance in percent).

See Also

chaid(), chaid_table()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
chaid_importance(fit)

Interactive CHAID tree plot with plotly

Description

Draws the tree as an interactive 'plotly' htmlwidget with zoom, pan and hover information (reaching rule, class distribution and split details for each node). The widget can be embedded directly in R Markdown documents or saved as HTML.

Usage

chaid_plotly(fit, palette = NULL, label_len = 20, ...)

Arguments

Value

A 'plotly' htmlwidget.

See Also

plot.chaid(), chaid_dot()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
chaid_plotly(fit)

Extract decision rules from a CHAID tree

Description

Returns the condition that a case must satisfy to reach each terminal node (or each of the requested nodes) as a rule string.

Usage

chaid_rules(fit, nodes = NULL, format = c("text", "sql", "r"))

Arguments

Value

A data frame with columns node (integer id) and rule (character).

See Also

chaid(), chaid_table()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
chaid_rules(fit)
chaid_rules(fit, format = "sql")

Summary table of CHAID terminal nodes

Description

Builds a segment summary of the terminal nodes. The first row is the root node, which serves as the baseline (index 100). For categorical responses the table contains the predicted class and the class shares, plus, when target is given, the response rate and the index value (response rate relative to the root, times 100). For continuous responses it contains the node mean, standard deviation and the index of the mean.

Usage

chaid_table(fit, target = NULL)

Arguments

Details

Note that n is the sum of frequency weights while pct_n and the class shares are computed with case weights included; the two scales coincide unless case weights are used.

Value

A data frame with one row for the root followed by one row per terminal node, including the reaching rule in the rule column.

See Also

chaid_rules(), chaid_gains(), chaid_importance()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
chaid_table(fit)
chaid_table(fit, target = "virginica")

Evaluate a CHAID tree on holdout data

Description

Routes newdata down the fitted tree and compares, for each terminal node, the node share and the response rate (node mean for continuous responses) between training and validation data. Useful to check whether the segments replicate on new data, i.e. whether the tree is overfitting.

Usage

chaid_validate(fit, newdata, weights = NULL, freq = NULL)

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

Arguments

Value

An object of class "chaid_validation": a list with nodes (per-node comparison data frame with train_pct_n, test_pct_n, train_rate, test_rate and diff_rate), overall (accuracy for categorical responses; RMSE and R-squared for continuous ones), n_test and ytype. A print() method is available.

See Also

chaid_gains(), predict.chaid()

Examples

set.seed(1)
idx <- sample(nrow(iris), 100)
fit <- chaid(Species ~ ., data = iris[idx, ],
             control = chaid_control(min_parent = 30, min_child = 10))
chaid_validate(fit, iris[-idx, ])

Plot a CHAID tree with base graphics

Description

Draws the tree top-down using base graphics only. For categorical responses each node shows the predicted class, its share, the node size and optionally a horizontal bar of the class distribution. Edges are labelled with the (possibly merged) predictor categories.

Usage

## S3 method for class 'chaid'
plot(
  x,
  cex = 0.8,
  label_len = 22,
  palette = NULL,
  show_bar = TRUE,
  main = NULL,
  ...
)

Arguments

Value

The fitted object, invisibly.

See Also

chaid(), chaid_dot() for publication-quality Graphviz output, chaid_plotly() for an interactive version.

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
plot(fit)

Predict from a fitted CHAID tree

Description

Routes the rows of newdata down the tree and returns predictions. Factor levels unseen during training, and codes that did not occur in a node when it was split, are routed to the child with the largest node size (for ordinal predictors, to the group with the smallest code distance); a warning is issued when unseen levels are detected.

Usage

## S3 method for class 'chaid'
predict(object, newdata, type = c("response", "prob", "node"), ...)

Arguments

Value

Depending on type: a factor (or numeric vector) of predictions, a numeric matrix of class probabilities with one column per response level, or an integer vector of node ids.

See Also

chaid()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
predict(fit, head(iris))
predict(fit, head(iris), type = "prob")
predict(fit, head(iris), type = "node")

Print and summarise a CHAID tree

Description

print() displays the tree structure with one line per node showing the prediction, node size and split information. summary() additionally reports the control settings and a risk estimate on the training data (misclassification rate or expected misclassification cost for categorical responses, weighted within-node variance for continuous responses).

Usage

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

## S3 method for class 'chaid'
summary(object, ...)

Arguments

Value

The fitted object, invisibly.

See Also

chaid()

Examples

fit <- chaid(Species ~ ., data = iris,
             control = chaid_control(min_parent = 30, min_child = 10))
print(fit)
summary(fit)