Skip to content

linkedin/isolation-forest

BranchesTags

Repository files navigation

Logo

isolation-forest

Build Status Release License

Table of contents

Introduction

This is a distributed Scala/Spark implementation of the Isolation Forest unsupervised outlier detection algorithm. It includes both the standard Isolation Forest and the Extended Isolation Forest, which uses random hyperplane splits to eliminate the axis-aligned bias of the original algorithm. The standard Isolation Forest also features support for ONNX export for easy cross-platform inference. This library was created by James Verbus from the LinkedIn Anti-Abuse AI team.

Features

  • Distributed training and scoring: The isolation-forest module supports distributed training and scoring in Scala using Spark data structures. It inherits from the Estimator and Model classes in Spark's ML library in order to take advantage of machinery such as Pipelines. Model persistence on HDFS is supported.
  • Extended Isolation Forest: The ExtendedIsolationForest variant uses random hyperplane splits instead of axis-aligned splits, eliminating the directional bias present in the standard algorithm. This is especially useful for detecting anomalies in data with correlated features or anomalies that don't align with individual feature axes.
  • Broad portability via ONNX: The isolation-forest-onnx module provides a Python-based converter to convert a trained standard IsolationForestModel to ONNX format for broad portability across platforms and languages. ONNX is an open format built to represent machine learning models. Note: ONNX export is currently supported for the standard IsolationForest only. The ExtendedIsolationForest uses hyperplane splits that are not compatible with the axis-aligned tree ensemble representation used by the ONNX converter.

Getting started

Building the library

To build using the default of Scala 2.13.14 and Spark 3.5.5, run the following:

./gradlew build

This will produce a jar file in the ./isolation-forest/build/libs/ directory.

If you want to use the library with arbitrary Spark and Scala versions, you can specify this when running the build command.

./gradlew build -PsparkVersion=3.5.5 -PscalaVersion=2.13.14

To force a rebuild of the library, you can use:

./gradlew clean build --no-build-cache

To just run the tests:

./gradlew test

Add an isolation-forest dependency to your project

Please check Maven Central for the latest artifact versions.

Gradle example

The artifacts are available in Maven Central, so you can specify the Maven Central repository in the top-level build.gradle file.

repositories {
    mavenCentral()
}

Add the isolation-forest dependency to the module-level build.gradle file. Here is an example for a recent spark scala version combination.

dependencies {
    implementation("com.linkedin.isolation-forest:isolation-forest_3.5.5_2.13:<latest-version>")
}

Maven example

If you are using the Maven Central repository, declare the isolation-forest dependency in your project's pom.xml file. Here is an example for a recent Spark/Scala version combination.

<dependency>
  <groupId>com.linkedin.isolation-forest</groupId>
  <artifactId>isolation-forest_3.5.5_2.13</artifactId>
  <version>&lt;latest-version&gt;</version>
</dependency>

Usage examples

Model parameters

Parameter Default Value Description
numEstimators 100 The number of trees in the ensemble.
maxSamples 256 The number of samples used to train each tree. If this value is between 0.0 and 1.0, then it is treated as a fraction. If it is >1.0, then it is treated as a count.
contamination 0.0 The fraction of outliers in the training data set. If this is set to 0.0, it speeds up the training and all predicted labels will be false. The model and outlier scores are otherwise unaffected by this parameter.
contaminationError 0.0 The error allowed when calculating the threshold required to achieve the specified contamination fraction. The default is 0.0, which forces an exact calculation of the threshold. The exact calculation is slow and can fail for large datasets. If there are issues with the exact calculation, a good choice for this parameter is often 1% of the specified contamination value.
maxFeatures 1.0 The number of features used to train each tree. If this value is between 0.0 and 1.0, then it is treated as a fraction. If it is >1.0, then it is treated as a count.
bootstrap false If true, draw sample for each tree with replacement. If false, do not sample with replacement.
randomSeed 1 The seed used for the random number generator.
featuresCol "features" The feature vector. This column must exist in the input DataFrame for training and scoring.
predictionCol "predictedLabel" The predicted label. This column is appended to the input DataFrame upon scoring.
scoreCol "outlierScore" The outlier score. This column is appended to the input DataFrame upon scoring.

Training and scoring

Here is an example demonstrating how to import the library, create a new IsolationForest instance, set the model hyperparameters, train the model, and then score the training data. data is a Spark DataFrame with a column named features that contains a org.apache.spark.ml.linalg.Vector of the attributes to use for training. In this example, the DataFrame data also has a label column; it is not used in the training process, but could be useful for model evaluation.

import com.linkedin.relevance.isolationforest._
import org.apache.spark.ml.feature.VectorAssembler

/**
  * Load and prepare data
  */

// Dataset from http://odds.cs.stonybrook.edu/shuttle-dataset/
val rawData = spark.read
  .format("csv")
  .option("comment", "#")
  .option("header", "false")
  .option("inferSchema", "true")
  .load("isolation-forest/src/test/resources/shuttle.csv")

val cols = rawData.columns
val labelCol = cols.last
 
val assembler = new VectorAssembler()
  .setInputCols(cols.slice(0, cols.length - 1))
  .setOutputCol("features")
val data = assembler
  .transform(rawData)
  .withColumnRenamed(labelCol, "label")
  .select("features", "label")

// scala> data.printSchema
// root
//  |-- features: vector (nullable = true)
//  |-- label: integer (nullable = true)

/**
  * Train the model
  */

val contamination = 0.1
val isolationForest = new IsolationForest()
  .setNumEstimators(100)
  .setBootstrap(false)
  .setMaxSamples(256)
  .setMaxFeatures(1.0)
  .setFeaturesCol("features")
  .setPredictionCol("predictedLabel")
  .setScoreCol("outlierScore")
  .setContamination(contamination)
  .setContaminationError(0.01 * contamination)
  .setRandomSeed(1)

val isolationForestModel = isolationForest.fit(data)
 
/**
  * Score the training data
  */

val dataWithScores = isolationForestModel.transform(data)

// scala> dataWithScores.printSchema
// root
//  |-- features: vector (nullable = true)
//  |-- label: integer (nullable = true)
//  |-- outlierScore: double (nullable = false)
//  |-- predictedLabel: double (nullable = false)

The output DataFrame, dataWithScores, is identical to the input data DataFrame but has two additional result columns appended with their names set via model parameters; in this case, these are named predictedLabel and outlierScore.

Saving and loading a trained model

Once you've trained an isolationForestModel instance as per the instructions above, you can use the following commands to save the model to HDFS and reload it as needed.

val path = "/user/testuser/isolationForestWriteTest"

/**
  * Persist the trained model on disk
  */

// You can ensure you don't overwrite an existing model by removing .overwrite from this command
isolationForestModel.write.overwrite.save(path)

/**
  * Load the saved model from disk
  */

val isolationForestModel2 = IsolationForestModel.load(path)

Extended Isolation Forest

The Extended Isolation Forest (EIF) generalizes the standard Isolation Forest by replacing axis-aligned splits with random hyperplane splits. This eliminates the bias that standard Isolation Forest has toward detecting anomalies along individual feature axes, making it more effective for data with correlated features or anomalies that lie along non-axis-aligned directions. For full details, see Hariri et al., "Extended Isolation Forest," 2018.

When to use Extended Isolation Forest

Use ExtendedIsolationForest instead of IsolationForest when:

  • Your data has correlated features where anomalies may not be separable along any single axis.
  • You observe "ghost" high-score regions along feature axes in standard IF that don't correspond to real anomalies (a known artifact of axis-aligned splits).
  • You want a more rotationally invariant anomaly detector.

The standard IsolationForest remains a good default for high-dimensional, uncorrelated data where axis-aligned splits are sufficient and computational cost matters.

Extended Isolation Forest parameters

ExtendedIsolationForest accepts all the same parameters as IsolationForest (see the model parameters table above), plus one additional parameter:

Parameter Default Value Description
extensionLevel numFeatures - 1 Controls the number of non-zero coordinates in each random hyperplane normal vector. extensionLevel + 1 coordinates are non-zero. 0 uses axis-aligned splits. The maximum value is numFeatures - 1 (fully extended), which is the default if not set. EIF stores these hyperplanes sparsely, so per-node storage and dot products scale with the number of non-zero coordinates rather than the full input dimension.

Important: interaction with maxFeatures. When maxFeatures < 1.0, each tree trains on a random subset of features. The extensionLevel is relative to this subspace, not the original dataset dimensionality. For example, if your data has 10 features and maxFeatures = 0.5, each tree uses 5 features, and the valid range for extensionLevel is [0, 4]. If not set, extensionLevel defaults to numFeatures - 1 for the resolved subspace. If you explicitly set extensionLevel to a value greater than numFeatures - 1, training will throw an error.

EIF hyperplanes are stored in the original feature space using only their non-zero coordinates. In practice, this means extensionLevel controls not only expressiveness but also node-local storage and traversal cost, while maxFeatures still determines the feature subspace available to each tree.

Extended Isolation Forest usage example

import com.linkedin.relevance.isolationforest.extended._

val contamination = 0.02
val extendedIsolationForest = new ExtendedIsolationForest()
  .setNumEstimators(100)
  .setBootstrap(false)
  .setMaxSamples(256)
  .setMaxFeatures(1.0)
  .setFeaturesCol("features")
  .setPredictionCol("predictedLabel")
  .setScoreCol("outlierScore")
  .setContamination(contamination)
  .setContaminationError(0.01 * contamination)
  .setExtensionLevel(5)  // Fully extended for a 6-feature dataset (extensionLevel + 1 = 6 non-zero coordinates)
  .setRandomSeed(1)

val extendedIsolationForestModel = extendedIsolationForest.fit(data)
val dataWithScores = extendedIsolationForestModel.transform(data)

Saving and loading works the same way as the standard model:

// Save
extendedIsolationForestModel.write.overwrite.save(path)

// Load
val loadedModel = ExtendedIsolationForestModel.load(path)

ONNX conversion for portable inference

Note: ONNX conversion is supported for the standard IsolationForestModel only. The ExtendedIsolationForestModel uses hyperplane-based splits that are not compatible with the axis-aligned tree ensemble representation used by the current ONNX converter.

Converting a trained model to ONNX

The artifacts associated with the isolation-forest-onnx module are available in PyPI.

The ONNX converter can be installed using pip. It is recommended to use the same version of the converter as the version of the isolation-forest library used to train the model.

pip install isolation-forest-onnx==<matching-version>

You can then import and use the converter in Python.

import os
from isolationforestonnx.isolation_forest_converter import IsolationForestConverter

# This is the same path used in the previous example showing how to save the model in Scala above.
path = '/user/testuser/isolationForestWriteTest'

# Get model data path
data_dir_path = path + '/data'
avro_model_file = os.listdir(data_dir_path)
model_file_path = data_dir_path + '/' + avro_model_file[0]

# Get model metadata file path
metadata_dir_path =  path + '/metadata'
metadata_file = os.listdir(path + '/metadata/')
metadata_file_path = metadata_dir_path + '/' + metadata_file[0]

# Convert the model to ONNX format (this will return the ONNX model in memory)
converter = IsolationForestConverter(model_file_path, metadata_file_path)
onnx_model = converter.convert()

# Convert and save the model in ONNX format (this will save the ONNX model to disk)
onnx_model_path = '/user/testuser/isolationForestWriteTest.onnx'
converter.convert_and_save(onnx_model_path)

Using the ONNX model for inference (example in Python)

import numpy as np
import onnx
from onnxruntime import InferenceSession

# `onnx_model_path` the same path used above in the convert and save operation
onnx_model_path = '/user/testuser/isolationForestWriteTest.onnx'
dataset_path = 'isolation-forest-onnx/test/resources/shuttle.csv'
num_examples_to_print = 10

# Load data
input_data = np.loadtxt(dataset_path, delimiter=',')
num_features = input_data.shape[1] - 1
last_col_index = num_features
print(f'Number of features: {num_features}')

# The last column is the label column
input_dict = {'features': np.delete(input_data, last_col_index, 1).astype(dtype=np.float32)}
actual_labels = input_data[:, last_col_index]

# Load the ONNX model from local disk and do inference
onx = onnx.load(onnx_model_path)
sess = InferenceSession(onx.SerializeToString())
res = sess.run(None, input_dict)

# Print scores
actual_outlier_scores = res[0]
print('ONNX Converter outlier scores:')
print(np.transpose(actual_outlier_scores[:num_examples_to_print])[0])

Performance and benchmarks

We benchmarked the standard Isolation Forest (StandardIF), Extended Isolation Forest at extension level 0 (ExtendedIF_0), and the fully extended variant (ExtendedIF_max) against the Liu et al. 2008 paper results and the reference Python EIF implementation (Ref. Python). All results use 100 trees, 256 samples per tree, 10 trials with unique random seeds, and report the mean ± standard error of the mean. The Ref. Python columns show EIF results at the corresponding extension level, not standard IF.

Dataset Dim Model AUROC AUPRC Liu et al. AUROC (IF) Ref. Python AUROC (EIF) Ref. Python AUPRC (EIF)
Annthyroid 6 StandardIF 0.813 ± 0.004 0.312 ± 0.004 0.82 - -
ExtendedIF_0 0.813 ± 0.004 0.307 ± 0.004 - 0.822 ± 0.004 0.314 ± 0.007
ExtendedIF_max 0.646 ± 0.002 0.1791 ± 0.0017 - 0.651 ± 0.003 0.183 ± 0.005
Arrhythmia 274 StandardIF 0.8064 ± 0.0019 0.494 ± 0.006 0.80 - -
ExtendedIF_0 0.802 ± 0.002 0.478 ± 0.004 - 0.796 ± 0.004 0.462 ± 0.005
ExtendedIF_max 0.810 ± 0.004 0.495 ± 0.005 - 0.803 ± 0.003 0.490 ± 0.004
Breastw 9 StandardIF 0.9864 ± 0.0003 0.9684 ± 0.0008 0.99 - -
ExtendedIF_0 0.9878 ± 0.0003 0.9726 ± 0.0008 - 0.9873 ± 0.0005 0.9704 ± 0.0016
ExtendedIF_max 0.9835 ± 0.0004 0.9569 ± 0.0015 - 0.9841 ± 0.0006 0.959 ± 0.002
Cardio 21 StandardIF 0.928 ± 0.002 0.565 ± 0.008 - - -
ExtendedIF_0 0.921 ± 0.002 0.553 ± 0.009 - 0.918 ± 0.003 0.546 ± 0.013
ExtendedIF_max 0.933 ± 0.002 0.541 ± 0.006 - 0.931 ± 0.002 0.547 ± 0.009
ForestCover 10 StandardIF 0.882 ± 0.006 0.051 ± 0.003 0.88 - -
ExtendedIF_0 0.865 ± 0.008 0.050 ± 0.005 - 0.872 ± 0.010 0.049 ± 0.004
ExtendedIF_max 0.688 ± 0.008 0.0138 ± 0.0003 - 0.662 ± 0.009 0.0129 ± 0.0004
Http (KDDCUP99) 3 StandardIF 0.99970 ± 0.00010 0.93 ± 0.02 1.00 - -
ExtendedIF_0 0.99410 ± 0.00010 0.392 ± 0.004 - 0.99390 ± 0.00010 0.379 ± 0.004
ExtendedIF_max 0.99410 ± 0.00010 0.379 ± 0.006 - 0.9939 ± 0.0003 0.371 ± 0.009
Ionosphere 33 StandardIF 0.8443 ± 0.0002 0.8014 ± 0.0003 0.85 - -
ExtendedIF_0 0.8568 ± 0.0006 0.8108 ± 0.0007 - 0.8556 ± 0.0016 0.808 ± 0.002
ExtendedIF_max 0.9075 ± 0.0002 0.8804 ± 0.0002 - 0.9061 ± 0.0014 0.876 ± 0.002
Mammography 6 StandardIF 0.8649 ± 0.0015 0.218 ± 0.007 0.86 - -
ExtendedIF_0 0.865 ± 0.002 0.220 ± 0.006 - 0.868 ± 0.002 0.229 ± 0.013
ExtendedIF_max 0.8630 ± 0.0010 0.190 ± 0.003 - 0.8639 ± 0.0016 0.184 ± 0.004
Mulcross 4 StandardIF 0.9910 ± 0.0009 0.852 ± 0.014 0.97 - -
ExtendedIF_0 0.938 ± 0.002 0.428 ± 0.009 - 0.960 ± 0.003 0.538 ± 0.017
ExtendedIF_max 0.940 ± 0.003 0.442 ± 0.011 - 0.941 ± 0.005 0.45 ± 0.02
Pima 8 StandardIF 0.668 ± 0.004 0.490 ± 0.003 0.67 - -
ExtendedIF_0 0.667 ± 0.004 0.507 ± 0.004 - 0.675 ± 0.005 0.514 ± 0.005
ExtendedIF_max 0.644 ± 0.003 0.498 ± 0.002 - 0.640 ± 0.004 0.493 ± 0.004
Satellite 36 StandardIF 0.717 ± 0.008 0.672 ± 0.008 0.71 - -
ExtendedIF_0 0.715 ± 0.004 0.675 ± 0.003 - 0.700 ± 0.004 0.664 ± 0.006
ExtendedIF_max 0.725 ± 0.003 0.704 ± 0.004 - 0.740 ± 0.005 0.711 ± 0.005
Shuttle 9 StandardIF 0.9971 ± 0.0002 0.9742 ± 0.0017 1.00 - -
ExtendedIF_0 0.9974 ± 0.0002 0.9789 ± 0.0014 - 0.99750 ± 0.00010 0.9805 ± 0.0010
ExtendedIF_max 0.9934 ± 0.0002 0.822 ± 0.004 - 0.9932 ± 0.0002 0.818 ± 0.003
Smtp (KDDCUP99) 3 StandardIF 0.9099 ± 0.0014 0.00450 ± 0.00010 0.88 - -
ExtendedIF_0 0.896 ± 0.002 0.00400 ± 0.00010 - 0.897 ± 0.002 0.00410 ± 0.00010
ExtendedIF_max 0.858 ± 0.003 0.0098 ± 0.0011 - 0.857 ± 0.003 0.014 ± 0.003

Key observations:

  • StandardIF results are in agreement with the original Liu et al. paper.
  • ExtendedIF_max closely matches the reference Python EIF across all 13 datasets.
  • EIF improves on high-dimensional datasets, e.g., ionosphere (AUROC 0.907 vs 0.844, AUPRC 0.880 vs 0.801) and satellite (AUROC 0.725 vs 0.717, AUPRC 0.704 vs 0.672). EIF underperforms IF on some datasets.
  • ExtendedIF_0 is not equivalent to StandardIF. Both use axis-aligned splits, but standard IF retries on constant features while EIF does not (matching the reference Python and C++ implementations). ExtendedIF_0 closely matches the reference Python EIF on all datasets.

Copyright and license

Copyright 2019 LinkedIn Corporation All Rights Reserved.

Licensed under the BSD 2-Clause License (the "License"). See License in the project root for license information.

Contributing

If you would like to contribute to this project, please review the instructions here.

Citing this project

If you use this library in your research or project, please cite it using the metadata in CITATION.cff, or use the following BibTeX entry:

@software{isolation_forest,
  author = {Verbus, James},
  title = {isolation-forest},
  year = {2019},
  url = {https://github.com/linkedin/isolation-forest},
  license = {BSD-2-Clause}
}

References

  • F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in 2008 Eighth IEEE International Conference on Data Mining, 2008, pp. 413–422.
  • F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation-based anomaly detection,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 6, no. 1, p. 3, 2012.
  • S. Hariri, M. Carrasco Kind, and R. J. Brunner, “Extended Isolation Forest,” IEEE Transactions on Knowledge and Data Engineering, 2019. DOI:10.1109/TKDE.2019.2947676, arXiv:1811.02141.
  • S. Hariri, "eif: Extended Isolation Forest for Anomaly Detection," https://github.com/sahandha/eif.
  • Shebuti Rayana (2016). ODDS Library [http://odds.cs.stonybrook.edu]. Stony Brook, NY: Stony Brook University, Department of Computer Science.

About

A distributed Spark/Scala implementation of the isolation forest algorithm for unsupervised outlier detection, featuring support for scalable training and ONNX export for easy cross-platform inference.

Topics

machine-learning scala spark linkedin outlier-detection unsupervised-learning anomaly-detection isolation-forest onnx

Resources

Readme

License

View license

Contributing

Contributing
Activity
Custom properties

Stars

252 stars

Watchers

13 watching

Forks

54 forks
Report repository

Releases 48

v4.1.1 Latest
Mar 19, 2026
+ 47 releases

Packages

 
 
 

Contributors 8

Languages