Alibi-detect

As explained in Outlier Detection in Python

An open-source Python library for detecting data drift, outliers, and other distributional or prediction anomalies across multiple data modalities.

alibi-detect is a Python library for outlier and drift detection across tabular, text, image, and time-series data. Outlier detection usually asks whether a single record is unusual; drift detection asks whether recent data has an unusual overall distribution compared with earlier data. Both can help monitor a deployed application and the data it processes. section 7.2

For tabular outliers, available methods include Isolation Forest and Mahalanobis distance. The former wraps scikit-learn’s implementation; the latter wraps a PyTorch method. Deep-learning options for image and tabular data include autoencoders, variational autoencoders, and Gaussian-mixture variants. AEGMM combines an autoencoder’s representation and reconstruction error with a Gaussian mixture model, training the components under a shared objective rather than fitting them separately. section 7.2 section 16.4

Practical considerations:

• Install the library before creating a detector. Its deep-learning use requires TensorFlow or PyTorch; installation extras include `alibi-detect[tensorflow]` and `alibi-detect[torch]`. section 16.4
• Expect more detector-construction code than with PyOD, but also more flexibility. Use the documentation and examples to understand a detector beyond a single worked listing. section 16.4
• Treat outlier and drift results as different signals: an increased proportion of outliers can indicate drift, but drift concerns the recent distribution rather than one record. section 7.2

Image reconstruction and perturbation

A CIFAR10 example normalizes images and constructs a convolutional `OutlierAE` detector. It compares an image with its reconstruction, then perturbs image values and predicts outliers with instance and feature scores. The illustrated unperturbed boat is reconstructed closely; after random noise is added, reconstruction error is high and the image is flagged. This demonstrates a detectable perturbation, not a guarantee that every flagged image has a meaningful problem. section 16.4

Listing 16.9 Evaluating the CIFAR10 image dataset

import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
tf.keras.backend.clear_session()
from tensorflow.keras.layers import Conv2D, Conv2DTranspose, \
    Dense, Layer, Reshape, InputLayer, Flatten
from alibi_detect.od import OutlierAE
 
train, test = tf.keras.datasets.cifar10.load_data()          #1
 
X_train, y_train = train
X_test, y_test = test
X_train = X_train.astype('float32') / 255
X_test = X_test.astype('float32') / 255
 
encoding_dim = 1024
 
encoder_net = tf.keras.Sequential([                          #2
      InputLayer(input_shape=(32, 32, 3)),
      Conv2D(64, 4, strides=2, padding='same', activation=tf.nn.relu),
      Conv2D(128, 4, strides=2, padding='same', activation=tf.nn.relu),
      Conv2D(512, 4, strides=2, padding='same', activation=tf.nn.relu),
      Flatten(),
      Dense(encoding_dim,)])
 
decoder_net = tf.keras.Sequential([                          #3
      InputLayer(input_shape=(encoding_dim,)),
      Dense(4*4*128),
      Reshape(target_shape=(4, 4, 128)),
      Conv2DTranspose(256, 4, strides=2, padding='same',
          activation=tf.nn.relu),
      Conv2DTranspose(64, 4, strides=2, padding='same', 
          activation=tf.nn.relu),
      Conv2DTranspose(3, 4, strides=2, padding='same', 
          activation='sigmoid')])
 
od = OutlierAE(threshold=.015,                               #4
               encoder_net=encoder_net,  
               decoder_net=decoder_net)
od.fit(X_train, epochs=5, verbose=True)
 
 
X = X_train[:500]                                            #5
od_preds = od.predict(X,
                      outlier_type='instance',    
                      return_feature_score=True,  
                      return_instance_score=True)
print("Number of outliers with normal data:",  
   od_preds['data']['is_outlier'].tolist().count(1))

The CIFAR10 listing shows image normalization and the start of a convolutional encoder for the autoencoder detector. section 16.4

Figure 16.8 The left pane shows image 8 with random noise added. The right pane shows the reconstruction of this. Here the reconstruction error is high, and the image in the left pane is flagged as an outlier.

The paired images make the high reconstruction error after perturbation visible. section 16.4

How Outlier Detection in Python covers it

How deep it goes

Explains detector families, wrappers, and AEGMM’s shared training objective, while directing detailed detector setup to documentation and examples.

Examples

Uses CIFAR10 images to compare autoencoder reconstructions before and after random perturbation.

What it assumes

Requires TensorFlow or PyTorch for deep-learning detectors and some detector-construction code; demonstrates setup through an image example.

What you can do after

Apply it via alibi-detect to construct an image autoencoder detector and inspect its reconstruction and outlier predictions.

How it connects

Connects outlier detection with data drift, Isolation Forest, Mahalanobis distance, autoencoders, variational autoencoders, and Gaussian mixture models.

liveBookmanning.com

Read Outlier Detection in Python on liveBook Get the book at manning.com

Related concepts

• PyTorchuses · PyTorch is an open-source machine-learning framework for Python that provides tensor computation, automatic differentiation, and tools for building and training neural networks.
• scikit-learnuses · scikit-learn is an open-source Python library providing tools for classical machine learning, including algorithms, preprocessing, model selection, and evaluation.
• TensorFlowuses · TensorFlow is an open-source software platform for numerical computation and machine learning, with tools for developing, training, and deploying models.
• data driftimplements · Data drift is a change over time in the statistical distribution or characteristics of data relative to a reference distribution or set of characteristics.
• outlier detectionuses · Outlier detection is the process of identifying observations that substantially deviate from an expected or prevailing pattern in data.
• mahalanobis distanceuses · Mahalanobis distance is a covariance-adjusted measure of separation between observations or between an observation and a reference distribution.

This explainer was generated by Manning's knowledge graph from the text of Outlier Detection in Python. The section links point to the parts of the book it draws on.