Support categorical features in ranger.unify and gbm.unify

[Copilot]

ranger.unify and gbm.unify rejected models trained on factor columns. This PR implements categorical split support across the full treeshap pipeline.

Encoding convention

A new decision_type = 3 (displayed as "==") represents a categorical bitmask split. The Split value is a right-group bitmask: bit k-1 set → factor level k goes to the No (right) child. Factor columns in datasets are passed through as.numeric(factor) (already done by existing sapply(x, as.numeric) calls), yielding 1-based integer codes.

C++ traversal (treeshap.cpp, set_reference_dataset.cpp, predict.cpp)

Added decision_type == 3 branch to all three tree-walking functions:

|| ((decision_type[j] == 3) && ((int)observation[feature[j]] >= 1)
    && !(((int)split[j] >> ((int)observation[feature[j]] - 1)) & 1))

If the observation's level is not in the right-group bitmask → goes left (Yes child).

gbm.unify

• Removed stop() that blocked categorical features.
• For each categorical split node, reads the 0-based c.splits index stored in Split, retrieves gbm_model$c.splits[[idx]], and computes a right-group bitmask (levels where c_split[k] == 1 go right). Sets Decision.type = "==" for those nodes.

ranger.unify / ranger_unify.common

• Extracts !forest$is.ordered (per-feature unordered flag) from the ranger model and passes it into ranger_unify.common.
• When treeInfo returns character splitval (partition mode), parses the comma-separated right-going level indices and computes the right-group bitmask. Threshold splits on ordered/numeric features in the same trees (also coerced to character by R) are handled by as.numeric() conversion.

model_unified.R

• is.model_unified() now accepts c("<=", "<", "==") as a valid level set (backward-compatible with existing two-level models).
• Updated Split and Decision.type column documentation.

Usage

# ranger with unordered factor splits (partition mode)
rf <- ranger::ranger(target ~ ., data = data_with_factors,
                     respect.unordered.factors = "partition")
unified <- ranger.unify(rf, data_with_factors)
treeshap(unified, data_with_factors[1:5, ])

# gbm with factor columns
gbm_model <- gbm::gbm(target ~ ., data = data_with_factors)
unified <- gbm.unify(gbm_model, data_with_factors)
treeshap(unified, data_with_factors[1:5, ])

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• cran.r-project.org
  • Triggering command: /home/REDACTED/work/_temp/ghcca-node/node/bin/node /home/REDACTED/work/_temp/ghcca-node/node/bin/node --enable-source-maps /home/REDACTED/work/_temp/copilot-developer-action-main/dist/index.js (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)