Prediction parity between LightGBM (Python) and leaves (Go)

[YuminosukeSato]

Summary

Under the same model and input, predictions from LightGBM's Python predict can differ from those produced by leaves. Root causes were specification mismatches and missing features in leaves:

• RF averaging mismatch: LightGBM averages by the total number of trees when average_output is set, while leaves averaged by the number of trees actually used.
• Post-transform coverage: Logistic/softmax/exponential post-transform was not applied for RF models and for JSON-loaded models, even when requested.
• Transform name gap: Exponential name was missing from TransformType.Name() which could lead to unknown.

Expected behavior

When Python's predict(X, raw_score=..., num_iteration=...) has a logical counterpart in leaves using
LGEnsembleFromFile(path, useTransformation) / LGEnsembleFromJSON(r, useTransformation) and Predict...(nEstimators), the numbers should match.

Actual behavior

• With RF models, using only a subset of trees (e.g., via num_iteration) caused a mismatch because of different averaging rules.
• Even with loadTransformation=true, RF and JSON paths did not apply logistic/softmax/exp transforms, so probabilities did not match Python.

Reproduction (example)

1. Python

clf = lgb.Booster(model_file='model.txt')
py = clf.predict(X, raw_score=False, num_iteration=k)

2. Go (before)

model, _ := leaves.LGEnsembleFromFile("model.txt", true)
_ = model.PredictDense(vals, rows, cols, pred, k, 0)

→ RF probabilities differ because the older implementation divided by used trees, not total trees.

For JSON models, LGEnsembleFromJSON did not apply the transform and returned raw scores.