Custom strategies

Step 1: C++ implementation

Each strategy family has a base class with pure virtual methods. Your new strategy inherits from the appropriate base and implements them.

For example, a new projection pursuit strategy needs to implement optimize() (find the best projection):

// File: core/src/models/strategies/pp/MyMethod.hpp
#pragma once
#include "models/strategies/pp/ProjectionPursuit.hpp"
#include "models/strategies/Strategy.hpp"
#include "utils/JsonValidation.hpp"

namespace ppforest2::pp {

struct MyMethod : public ProjectionPursuit {
  explicit MyMethod(float alpha) : alpha_(alpha) {}

  std::string display_name() const override { return "My method"; }

  Result optimize(
    const types::FeatureMatrix&  x,
    const stats::GroupPartition& group_spec) const override {
    // Find the optimal projector for the data.
    // Return Result{ projector_vector, index_value }.
    ...
  }

  nlohmann::json to_json() const override {
    return {{"name", "my_method"}, {"alpha", alpha_}};
  }

  static ProjectionPursuit::Ptr from_json(const nlohmann::json& j) {
    JsonReader{j, "my_method"}.only_keys({"name", "alpha"});
    return my_method(j.at("alpha").get<float>());
  }

  PPFOREST2_REGISTER_STRATEGY(ProjectionPursuit, "my_method")

private:
  const float alpha_;
};

inline ProjectionPursuit::Ptr my_method(float alpha) {
  return std::make_shared<MyMethod>(alpha);
}

}  // namespace ppforest2::pp

The key pieces:

• to_json() serializes the strategy name and parameters. This is used for model persistence.
• from_json() deserializes from JSON and validates that no unexpected keys are present.
• PPFOREST2_REGISTER_STRATEGY registers the factory so JSON deserialization finds it automatically.
• display_name() returns a human-readable label for summaries.
• Factory function (my_method()) is a convenience wrapper.

The same pattern applies to variable selection strategies (select()), cutpoint strategies (cutpoint()), and the other strategy families. See the C++ documentation for complete interface definitions and examples.

After writing the .cpp file, add it to core/src/models/CMakeLists.txt.

Step 2: R constructor

Write an R function that creates a strategy list. The name field must match the C++ registration name exactly.

#' My custom projection pursuit strategy.
#'
#' @param alpha A tuning parameter.
#' @return A \code{pp_strategy} object.
#' @export
pp_my_method <- function(alpha = 1.0) {
  if (!is.numeric(alpha) || length(alpha) != 1)
    stop("`alpha` must be a single number.")

  structure(
    list(name = "my_method", display_name = "My method", alpha = alpha),
    class = "pp_strategy"
  )
}

The constructor should:

• Validate parameters before they reach C++. Catching errors early with clear messages is better than a C++ exception.
• Set the S3 class to pp_strategy, vars_strategy, cutpoint_strategy, stop_strategy, binarize_strategy, or grouping_strategy. This is checked by resolve_strategies().
• Include display_name for readable output in summary().
• Use the same parameter names as to_json() in C++. The R list is converted to JSON and passed to from_json() on the C++ side.

Step 3: Use it

Once both sides are in place, the new strategy works like any built-in:

# Single tree
tree <- pptr(Species ~ ., data = iris, pp = pp_my_method(alpha = 0.5))

# Forest
forest <- pprf(Species ~ ., data = iris, pp = pp_my_method(alpha = 0.5), vars = vars_uniform(n_vars = 2))

# Summary shows the strategy
summary(tree)

The strategy is also available from the CLI:

ppforest2 train -d iris.csv --pp my_method:alpha=0.5

And models trained with the new strategy can be saved and loaded as usual — the JSON registry handles serialization automatically.

PP: Projection pursuit

Controls how the tree finds the best linear combination of variables at each node.

index(x, group_spec, projector) -> scalar
optimize(x, group_spec) -> Result{projector, index}

optimize() is the main method. It receives the data matrix and group partition and returns the best projection vector. index() evaluates a given projection (used for variable importance calculations).

Vars: Variable selection

Controls which variables are available to projection pursuit at each split. This is what makes random forests “random”.

select(x, group_spec, rng) -> Result{selected_indices, original_cols}

The returned VariableSelection::Result tracks which columns were selected so the reduced-space projector can be expanded back to the full feature space.

Threshold: Split cutpoint

Controls where the split cutpoint is placed in the projected space.

cutpoint(group_1, group_2, projector) -> scalar

Receives the two groups (already partitioned by projection pursuit) and the projection vector. Returns the cutpoint value.

Stop: Stopping rule

Controls when to stop growing the tree.

should_stop(group_partition, depth) -> bool

Binarize: Binarization

Controls how multiclass nodes (>2 groups) are reduced to a binary problem.

regroup(projected_x, group_partition) -> Result

Grouping: Group partition

Controls how observations are routed to children after a split.

split(partition, lower_group, upper_group) -> Result