pyadts 0.0.1b0

A python package for time series anomaly detection

Anomaly Detection for Time-series (PyADTS)

PyADTS is aimed at accelerating the workflow of time series anomaly detection for researchers. It contains various utilities for data loading, pre-processing, detector construction, detector ensemble, evaluation and etc. PyADTS can help you to write less boilerplate on following parts:

• Preparing dataset & pre-processing
• Feature extraction (Optional)
• Model training
• Ensemble (Optional)
• Evaluation

You can find the complete documentation here.

Table of Contents:

• Anomaly Detection for Time-series (PyADTS)
  • Installation
  • Quick Start
  • Other Utilities
  • Implemented Algorithms

Installation

To install the package locally, run:

>> cd <pyadts_dir>
>> pip install .

To install the package from PyPi, run:

>> pip install pyadts

Quick Start

Fetch the dataset

PyADTS contains various built-in datasets. To utilize them:

from pyadts.data import get_nab_nyc_taxi

data = get_nab_nyc_taxi(root_path='<the_path_of_nab_dataset>')

All components of the dataset are organized as a dict:

{'value': value, 'label': label, 'timestamp': timestamp, 'datetime': datetime}

Pre-processing

It's important to pre-process the time series before training. PyADTS offered three types of pre-processing methods including:

• Rearrangement: Sort the values along with the timestamp and reconstruct the timestamp to discover missing values. (return a dict and append an attribute missing)
• Normalization: Normalize the time series
• Imputation: Impute the time series.

from pyadts.data import series_impute, series_normalize, series_rearrange

data_processed = series_rearrange(**data)

data_processed['value'] = series_normalize(data_processed['value'], mask=data_processed['missing'], method='zscore')

data_processed['value'] = series_impute(data_processed['value'], missing=data_processed['missing'], method='linear')

Feature extraction

Extracting manual features is essential for some anomaly detection approaches. PyADT offered various options for extracting features including:

• Simple features: logarithm, difference, second-order difference, ...
• Window-based features: window mean value, window std value, ...
• Decomposition features: STL decomposition, ...
• Frequency domain features: wavelet features, spectral residual, ...
• Regression features: SARIMA regression residual, Exponential Smoothing residual, ...

from pyadts.data import FeatureExtractor

feature_extractor = FeatureExtractor()

Train the model

Different anomaly detection algorithms should be utilized to tackle different scenarios. PyADT contains various algorithms including supervised-, unsupervised-, nonparametric-methods (you can refer the full list of implemented algorithms).

from pyadts import ThresholdDetector

train_x = data['value']
detector = ThresholdDetector()
pred_y = detector.fit_predict(train_x)

Ensemble

TODO

Evaluation

It's easy to evaluate your algorithms using PyADT's built-in metrics:

from pyadts import roc_auc

train_y = data['label']
roc = roc_auc(pred_y, train_y, delay=7)

In real-world applications, the delay of anomaly alerts is acceptable. So PyADT offered the delay argument for all metrics.

The pipeline

TODO

Other Utilities

Visualization

You can visualize your data with a single line of code:

from pyadts.data import plot_series

fig = plot_series(value=data['value'], label=data['label'], datetime=data['datetime'], plot_vline=True)
fig.show()

The example visualization:

Generate synthetic data

TODO

Implemented Algorithms

Supervised Approaches

• Random Forest
• SVM
• Deep Neural Network

Unsupervised Approaches

Non-parametric

• SR
• Threshold
• Quantile
• K​-Sigma

Statistic-based

• SPOT
• DSPOT
• Autoregression
• ESD
• S-ESD
• S-H-ESD

Machine learning-based

• LOF
• Isolation Forest
• OCSVM

Deep-based

• Autoencoder
• RNN Autoencoder
• Donut

TODO

• Fix bugs and reformat code
• Multi-variate time series support
• Complete the models
• Incorporating Numba
• Implement the pipeline
• Synthetic data generator
• Allow both inplace and non-inplace operations
• Add streaming support