feat: Implement new xgboost by gandola · Pull Request #156 · feedzai/feedzai-openml-java

gandola · 2026-09-18T20:57:06Z

Adds a new OpenML provider, openml-xgboost, built directly on the upstream XGBoost JVM binding (ml.dmlc:xgboost4j_2.12:3.4.0) to be a replacement of the H2O-embedded XGBoost.

Phase 1 covers:

Model import — loadModel / loadSchema / validateForLoad.
Real-time single-instance scoring — classify / getClassDistribution via Booster.inplace_predict.
In-process (non-Spark) training — fit via XGBoost.train.

Key classes

XgboostModelProvider — TrainingMachineLearningProvider, algorithm XGBOOST_BINARY_CLASSIFIER.
XgboostModelCreator — loadModel / fit / validateForLoad / validateForFit.
XgboostClassificationModel — scoring; native handle guarded by a synchronized block (the Booster handle is not thread-safe).
XgboostSchemaUtils — shared feature-row builder used by both training and scoring (guarantees identical encoding).

Native runtime

xgboost4j bundles native libs for linux/{x86_64,aarch64}, macos/{x86_64,aarch64}, windows/x86_64 — ARM/Graviton works out of the box.
Requires OpenMP (libgomp/libomp) at runtime.
No musl/Alpine native binary is published — native-dependent tests are Assume-guarded and skip gracefully there.

Dependency handling (Kryo)

xgboost4j_2.12 transitively pulls Kryo 5, whose com.esotericsoftware.kryo.Kryo collides on the classpath with the Kryo 4 (kryo-shaded) used by Spark/consumers (Kryo 5 removed the nested Kryo.DefaultInstantiatorStrategy). Since Booster implements com.esotericsoftware.kryo.KryoSerializable, kryo must still be
present at class-load. Resolution: exclude kryo from the xgboost4j dependency and re-declare it in provided scope — available for this module's own compile/tests, not propagated transitively to consumers (which supply their own Kryo).

Testing

20 tests; native-dependent ones Assume-skip where the native lib can't load (musl).
Coverage: 154/159 lines, 38/40 branches. Remaining gaps are the jackson-backed loadSchema one-line delegation and two native-exception catch blocks.
Verified train + save/reload produces identical scores; validation rejects string/free-text feature fields.

codecov · 2026-09-21T09:55:37Z

Codecov Report

❌ Patch coverage is 93.71069% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.74%. Comparing base (18a41e9) to head (96dad3b).

Files with missing lines	Patch %	Lines
...l/provider/xgboost/XgboostClassificationModel.java	83.33%	5 Missing and 1 partial ⚠️
...i/openml/provider/xgboost/XgboostModelCreator.java	95.40%	3 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##             master     #156      +/-   ##
============================================
+ Coverage     81.76%   82.74%   +0.98%     
- Complexity      525      579      +54     
============================================
  Files            52       58       +6     
  Lines          1771     1930     +159     
  Branches        179      193      +14     
============================================
+ Hits           1448     1597     +149     
- Misses          234      242       +8     
- Partials         89       91       +2

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

fdz-joao-santos · 2026-09-21T13:29:06Z

+/**
+ * Specifies the XGBoost algorithms that can be imported and trained through this provider.
+ *
+ * @since 1.0.0


Not true, also, not needed.

jpleitao

Overall the code looks good, just added some comments regarding (i) the set of parameters being exposed, (ii) default values for some parameters, and (iii) validations

jpleitao · 2026-10-13T21:58:53Z

+    XGBOOST_BINARY_CLASSIFIER(createDescriptor(
+            "DMLC - XGBoost",
+            XgboostDescriptorUtil.PARAMS,
+            MachineLearningAlgorithmType.SUPERVISED_BINARY_CLASSIFICATION,


See also the comment in validateForFit(), but in short XGBoost is being declared here as binary classification, but no such validation that the target is actually binary is being performed

jpleitao · 2026-10-13T22:06:11Z

+    /**
+     * The set of parameters accepted when training an XGBoost model.
+     */
+    static final Set<ModelParameter> PARAMS = ImmutableSet.of(


Looking at the XGBoost support in an internal data science Python package, and the XGBoost documentation, it seems to me that some supported parameters are note defined here:

max_delta_step (default 0.0) — Maximum absolute value of a leaf node prediction

colsample_bylevel (default 1.0) — Column sample rate

tree_method (default auto/hist) - Tree construction algorithm used in XGBoost. Possible values: auto, exact, approx, hist, where auto is the same as hist

grow_policy (default depthwise) - Controls a way new nodes are added to the tree. Possible values: depthwise, lossguide

max_leaves (default 0) — Maximum number of nodes to be added. Not used by exact tree method.

max_bin (default 256) — Maximum number of bins. Only used if tree_method is set to hist or approx

While not defined in the internal DS Python package, could also make sense to define:

scale_pos_weight (default 1.0) — Control the balance of positive and negative weights, useful for unbalanced classes.

jpleitao · 2026-10-13T22:08:59Z

+            new ModelParameter(
+                    OBJECTIVE_PARAMETER_NAME,
+                    "Objective",
+                    "The learning task and corresponding objective:\n"


If XGBoost is declared to be MachineLearningAlgorithmType.SUPERVISED_BINARY_CLASSIFICATION, then it doesn't make sense to allow for objective = multi:softprob (it also requires a num_class parameter indicating the number of classes, but since this XGboost is for binary classification, no such parameter is exposed).

My suggestion is to remove multi:softprob from the set of possible values of this parameter and add a check in XgboostModelCreator.validateForFit() thorwing an error in case objective=multi:softprob is specified

jpleitao · 2026-10-13T22:12:44Z

+                    "Objective",
+                    "The learning task and corresponding objective:\n"
+                            + "'binary:logistic' outputs the probability of the positive class,\n"
+                            + "'binary:logitraw' outputs the raw (pre-sigmoid) score,\n"


objective = binary:logitraw outputs a raw pre-sigmoid score (unbounded, can be negative), but XgboostClassificationModel.toClassDistribution() treats any single-value output as a probability and returns [1-p, p]. As such, when objective = binary:logitraw is specified, [1-p, p] can fall outside the [0, 1] interval.

My suggestion is to remove binary:logitraw from the set of possible values for this parameter, and add a check in XgboostModelCreator.validateForFit() throwing an error in case objective=binary:logitraw is specified. It is also possible to keep allowing binary:logitraw and apply a sigmoid in XgboostModelCreator.validateForFit(), but that would produce the same as using objective=binary:logistic

jpleitao · 2026-10-13T22:13:46Z

+                    "Number of boosting rounds",
+                    "Number of boosting iterations (trees) to build.",
+                    MANDATORY,
+                    intRange(1, Integer.MAX_VALUE, 100)


In the internal DS Python package a default value of 50 is being used. Consider adopting the same default value here in order to align with the previous implementation

jpleitao · 2026-10-13T22:16:37Z

+                    "Number of threads",
+                    "Number of parallel threads used to run XGBoost. Defaults to 1 for deterministic behavior.",
+                    NOT_MANDATORY,
+                    intRange(1, Integer.MAX_VALUE, 1)


In the internal DS Python package, this parameter takes None as default value, which means all available threads can be used. In native XGBoost, this is achieved by leaving this unset or specifying nthread=0, so I think the minimum value of this parameter should be changed to 0. The description of the parameter could also be updated in order to specify that 0 means using all available threads.

It is also worth noting that, while keeping 1 as the default value is safe, the default in H2O was to use all threads, meaning that this code could introduce slower default training.

jpleitao · 2026-10-13T22:21:21Z

+
+            booster = XGBoost.train(trainMatrix, boosterParams, numRound, new HashMap<>(), null, null);
+
+            final Path exportDir = exportModel(booster);


Could be worth adding a comment explaining the rationale being saving the model to file and reading again: my understanding is that something similar is already done in this repository for LightGBM, and we are also ensuring that fit() shares the same code as model import, so a freshly-trained model scores identically to the same model saved and reloaded.

But I could be missing or misunderstanding something.

jpleitao · 2026-10-13T22:24:13Z

+            booster = XGBoost.train(trainMatrix, boosterParams, numRound, new HashMap<>(), null, null);
+
+            final Path exportDir = exportModel(booster);
+            return loadModel(exportDir, schema);


In exportModel() a temporary directory is created but after loadModel() it is never removed.

Consider clearing that temporary folder in the finally block. As a starting point, this is what is currently being done for LightGBM (single temporary file instead of temporary directory):

try { Files.delete(tmpModelFilePath); } catch (final IOException e) { logger.error("Could not delete temporary model file: {}", e.getMessage()); }

jpleitao · 2026-10-13T22:25:05Z

+    @Override
+    public List<ParamValidationError> validateForFit(final Path pathToPersist,
+                                                     final DatasetSchema schema,
+                                                     final Map<String, String> params) {


Not sure if this is the best place to this, but could be worth performing some validations that the target is binary, just like it is done for LightGBM:

if (getNumTargetClasses(schema).orElse(-1) != 2) { errorsBuilder.add(new ParamValidationError(ERROR_MSG_NON_BINARY_TARGET)); }

gandola added 2 commits September 18, 2026 21:49

feat: Implement new xgboost

8bf74cd

feat: Fix some cicd issues

7814168

fdz-joao-santos reviewed Sep 21, 2026

View reviewed changes

feat: Change the name of the algorithm

96dad3b

gandola requested a review from a team October 13, 2026 18:00

jpleitao suggested changes Oct 13, 2026

View reviewed changes

Navigation Menu

Repository navigation

feat: Implement new xgboost - #156

feat: Implement new xgboost#156
gandola wants to merge 3 commits into
masterfrom
ft-pg-RLAB-617

gandola commented Sep 18, 2026 •

edited

Loading

Uh oh!

codecov Bot commented Sep 21, 2026 •

edited

Loading

Uh oh!

fdz-joao-santos Sep 21, 2026 •

edited

Loading

Uh oh!

jpleitao left a comment

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

jpleitao Oct 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants


		booster = XGBoost.train(trainMatrix, boosterParams, numRound, new HashMap<>(), null, null);

		final Path exportDir = exportModel(booster);

Conversation

gandola commented Sep 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Sep 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

fdz-joao-santos Sep 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jpleitao left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

gandola commented Sep 18, 2026 •

edited

Loading

codecov Bot commented Sep 21, 2026 •

edited

Loading

fdz-joao-santos Sep 21, 2026 •

edited

Loading