Implementation:Online ml River Anomaly OneClassSVM

Knowledge Sources	River River Docs Estimating the Support of a High-Dimensional Distribution
Domains	Online Machine Learning, Anomaly Detection, Support Vector Machines
Last Updated	2026-02-08 16:00 GMT

Overview

Concrete tool for performing online one-class SVM anomaly detection in the River library, implementing stochastic gradient descent optimization of the nu-SVM objective for streaming data.

Description

The anomaly.OneClassSVM class implements an online variant of the One-Class SVM algorithm for unsupervised anomaly detection. It inherits from both linear_model.base.GLM (for the SGD machinery) and anomaly.base.AnomalyDetector (for the anomaly detection interface).

Internally, the model treats every observation as a positive example (y=1) and optimizes a hinge loss with L2 regularization controlled by nu / 2. The decision function score_one(x) returns w . x - intercept, where lower values indicate more anomalous observations.

It is recommended to:

Scale data upstream with preprocessing.StandardScaler
Use feature_extraction.RBFSampler to capture non-linearities
Wrap with anomaly.QuantileFilter to convert raw scores to binary anomaly labels

Usage

Import and use anomaly.OneClassSVM when:

You need an SGD-based online anomaly detector
You want to leverage the SVM framework for unsupervised anomaly detection
You can estimate the expected anomaly fraction (nu parameter)
You want raw decision function scores for flexible downstream thresholding

Code Reference

Source Location

river/anomaly/svm.py, lines 8-111.

Signature

class OneClassSVM(linear_model.base.GLM, anomaly.base.AnomalyDetector):
    def __init__(
        self,
        nu=0.1,
        optimizer: optim.base.Optimizer | None = None,
        intercept_lr: optim.base.Scheduler | float = 0.01,
        clip_gradient=1e12,
        initializer: optim.base.Initializer | None = None,
    ):

Import

from river import anomaly
model = anomaly.OneClassSVM(nu=0.1)

Parameters

Parameter	Type	Default	Description
nu	float	0.1	Upper bound on the fraction of training errors and lower bound on the fraction of support vectors. Interpreted as the expected fraction of anomalies.
optimizer	optim.base.Optimizer or None	None	Sequential optimizer for weight updates. Defaults to `optim.SGD(0.01)`.
intercept_lr	float or Scheduler	0.01	Learning rate for the intercept. Set to 0 to disable intercept updates.
clip_gradient	float	1e12	Clips the absolute value of each gradient value.
initializer	optim.base.Initializer or None	None	Weight initialization scheme. Defaults to `optim.initializers.Zeros()`.

Methods

learn_one(x: dict) -> None -- Updates the model weights using SGD on the hinge loss, treating x as a positive (normal) example.
score_one(x: dict) -> float -- Returns the raw decision function value w . x - intercept. Lower values indicate more anomalous observations.

I/O Contract

Inputs

Method	Parameter	Type	Description
learn_one	x	dict	A dictionary mapping feature names to numeric values.
score_one	x	dict	A dictionary mapping feature names to numeric values.

Outputs

Method	Return Type	Description
learn_one	None	Updates internal weights via SGD; no return value.
score_one	float	Raw decision function value. Not bounded to [0, 1]. Lower values indicate more anomalous observations.

Usage Examples

Basic usage with QuantileFilter:

from river import anomaly, datasets, metrics

model = anomaly.QuantileFilter(
    anomaly.OneClassSVM(nu=0.2),
    q=0.995
)

auc = metrics.ROCAUC()

for x, y in datasets.CreditCard().take(2500):
    score = model.score_one(x)
    is_anomaly = model.classify(score)
    model.learn_one(x)
    auc.update(y, is_anomaly)

print(auc)
# ROCAUC: 74.68%

Evaluation with progressive_val_score:

from river import anomaly, datasets, metrics, evaluate

model = anomaly.QuantileFilter(
    anomaly.OneClassSVM(nu=0.2),
    q=0.995
)

evaluate.progressive_val_score(
    dataset=datasets.CreditCard().take(2500),
    model=model,
    metric=metrics.ROCAUC(),
    print_every=1000
)
# [1,000] ROCAUC: 74.40%
# [2,000] ROCAUC: 74.60%
# [2,500] ROCAUC: 74.68%
# ROCAUC: 74.68%

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment