Maintenance notice: This JuliaHEP fork is being retired and will be archived. Use EvoTrees.jl for new work. This repository and its tutorials remain available as reference material; further development is not planned.
The comparison in issue #8 found matching controlled logistic calculations, broader EvoTrees functionality, and substantially faster EvoTrees training in the measured workloads. The report, scripts, and raw results preserve the evidence and reproduction instructions. JLBoost's exact split search and explicit, easily reweighted trees remain useful for study, but do not justify maintaining a separate general-purpose boosting library for this fork's needs.
A 100%-Julia implementation of gradient boosting regression trees (GBRT / GBDT).
This is the JuliaHEP fork of xiaodaigh/JLBoost.jl. It requires Julia 1.10+.
The remaining usage examples and longer tutorials in docs/ are retained for reference.
using Pkg
Pkg.add(url="https://github.com/JuliaHEP/JLBoost.jl")jlboost fits on any Tables.jl table. The default loss is logistic (LogitLogLoss) for a binary target. Features default to every column except the target.
using JLBoost, DataFrames
df = DataFrame(
x = Float64[1, 1, 1, 1, 0, 0, 0, 0],
y = [1, 1, 1, 1, 0, 0, 0, 0],
)
model = jlboost(df, :y; nrounds = 2, max_depth = 2)
ŷ = predict(model, df) # raw margin
ŷ = model(df) # same: the model is callableA fitted JLBoostTreeModel holds the trees, the loss, and the target name:
trees(model)
model.loss # LogitLogLoss()
model.target # :yControl boosting with keywords on jlboost:
jlboost(df, :y;
nrounds = 10,
max_depth = 3, # depth-wise growth (set `0` if you use `max_leaves`)
max_leaves = 0, # leaf-wise / best-first growth when > 0
eta = 0.3, # shrinkage; kept on each tree, not baked into leaves
lambda = 0, # L2 on leaf scores
gamma = 0, # minimum gain to split
subsample = 1,
min_child_weight = 1,
weights = nothing, # optional per-row observation weights
)Score several models by concatenating them:
predict(vcat(model, model), df)Pass a vector the same length as the table. Gradients and hessians are multiplied by w_i.
w = fill(2.0, nrow(df))
model_w = jlboost(df, :y; weights = w)Leaf scores live in tree.weight. Shrinkage is a wrapper (WeightedJLBoostTree.eta), so you can rescale a tree after fitting:
t = trees(model)[1]
t.eta
0.3 * t # new tree with 30% of the original eta
feature_importance(model, df) # Quality_Gain, Coverage, FrequencySwap in a LossFunctions.jl loss. Least squares is L2DistLoss():
using LossFunctions: L2DistLoss
df = DataFrame(x = rand(100) .* 100)
df.y = 2 .* df.x .+ rand(100)
jlboost(df, :y, [:x], fill(0.0, nrow(df)), L2DistLoss(); max_depth = 2)JLBoost.save(model, "model.jlb")
JLBoost.load("model.jlb")Any column-accessible table works. For a custom type, define efficient nrow, ncol, and view if the generic methods are too slow.
| docs/ | Reference tutorials |
| Classification | Binary logistic example, AUC / gini, leaf-wise growth |
| Out of core | Fit from a JDF.JDFFile without loading every column |
| Give Me Some Credit | Larger example on disk |
| Λc classification | LHCb-like signal vs background, ROC vs a 1D Lc_PT cut |
- Julia 1.10+ (tested on 1.10, 1.11, and latest 1.x).
- Per-row observation
weights:jlboost(df, target; weights = w)(XGBoostDMatrixweight). - Training-set raw margins agree with XGBoost to ~10⁻⁸ under matched exact-split logistic settings. The 8-row stump is bit-identical.
- GitHub Actions tests: Julia 1.11 and latest 1.x on Linux, macOS, and Windows (push); Ubuntu + latest 1.x on pull requests.
Union{T, Missing}features are not supported yet.- Only a single target column (no multivariate models yet).
- Numeric and boolean features only (no categoricals yet).
There is no MLJ interface in this fork. The old JLBoostMLJ.jl wrapper is archived (last commit 2021) and depends on JLBoost 0.1, DataFrames 0.21, and MLJ 0.10–0.14, so it will not resolve against this package.