Anomaly Detection on Time Series: An Evaluation of Deep Learning Methods.

The goal of this repository is to provide a benchmarking pipeline for anomaly detection on time series data for multiple state-of-the-art deep learning methods.

Implemented Algorithms

LSTM-AD

Malhotra, Pankaj, et al. "Long short term memory networks for anomaly detection in time series." Proceedings. Presses universitaires de Louvain, 2015.

LSTM-ED

Malhotra, Pankaj, et al. "LSTM-based encoder-decoder for multi-sensor anomaly detection." ICML, 2016.

Autoencoder

Hawkins, Simon, et al. "Outlier detection using replicator neural networks." DaWaK, 2002.

Donut

Xu, Haowen, et al. "Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications." WWW, 2018.

REBM using a restricted Boltzmann Machine as energy-based model

Zhai, Shuangfei, et al. "Deep structured energy based models for anomaly detection." ICML, 2016.

DAGMM

Zong, Bo, et al. "Deep autoencoding gaussian mixture model for unsupervised anomaly detection." ICLR, 2018.

LSTM-DAGMM

Extension of Dagmm using an LSTM-Autoencoder instead of a Neural Network Autoencoder

Usage

git clone git://github.com/KDD-OpenSource/DeepADoTS.git  
virtualenv venv -p /usr/bin/python3  
source venv/bin/activate  
pip install -r requirements.txt  
python3 main.py

Example

We follow the scikit-learn API by offering the interface methods fit(X) and predict(X). The former estimates the data distribution in an unsupervised way while the latter returns an anomaly score for each instance - the higher, the more certain is the model that the instance is an anomaly. To compare the performance of methods, we use the ROC AUC value.

We use MNIST to demonstrate the usage of a model since it is already available in TensorFlow and does not require downloading external data (even though the data has no temporal aspect).

import pandas as pd
import tensorflow as tf
from sklearn.metrics import roc_auc_score

from src.algorithms import AutoEncoder
from src.datasets import Dataset


class MNIST(Dataset):
    """0 is the outlier class. The training set is free of outliers."""

    def __init__(self, seed):
        super().__init__(name="MNIST", file_name=None)  # We do not need to load data from a file
        self.seed = seed

    def load(self):
        # 0 is the outlier, all other digits are normal
        OUTLIER_CLASS = 0
        mnist = tf.keras.datasets.mnist
        (x_train, y_train), (x_test, y_test) = mnist.load_data()
        # Label outliers with 1 and normal digits with 0
        y_train, y_test = (y_train == OUTLIER_CLASS), (y_test == OUTLIER_CLASS)
        x_train = x_train[~y_train]  # Remove outliers from the training set
        x_train, x_test = x_train / 255, x_test / 255
        x_train, x_test = x_train.reshape(-1, 784), x_test.reshape(-1, 784)
        self._data = tuple(pd.DataFrame(data=data) for data in [x_train, y_train, x_test, y_test])


x_train, y_train, x_test, y_test = MNIST(seed=0).data()
# Use fewer instances for demonstration purposes
x_train, y_train = x_train[:1000], y_train[:1000]
x_test, y_test = x_test[:100], y_test[:100]

model = AutoEncoder(sequence_length=1, num_epochs=40, hidden_size=10, lr=1e-4)
model.fit(x_train)

error = model.predict(x_test)
print(roc_auc_score(y_test, error))  # e.g. 0.8614

We can visualize the samples with respective error values as follows

import numpy as np
import matplotlib.pyplot as plt
from matplotlib import offsetbox

"""Borrowed from https://github.com/scikit-learn/scikit-learn/blob/master/examples/manifold/plot_lle_digits.py#L44"""
error = (error - error.min()) / (error.max() - error.min())  # Normalize error
x_test = x_test.values
y_random = np.random.rand(len(x_test)) * 2 - 1
plt.figure(figsize=(20, 10))
ax = plt.subplot(111)
if hasattr(offsetbox, 'AnnotationBbox'):
    shown_images = np.array([[1., 1.]])
    for i in range(len(x_test)):
        X_instance = [error[i], y_random[i]]
        dist = np.sum((X_instance - shown_images) ** 2, 1)
        if np.min(dist) < 4e-5:
            # don't show points that are too close
            continue
        shown_images = np.r_[shown_images, [X_instance]]
        imagebox = offsetbox.AnnotationBbox(offsetbox.OffsetImage(x_test[i].reshape(28, 28), cmap=plt.cm.gray_r), X_instance)
        ax.add_artist(imagebox)
plt.xlim((0, 1.1))
plt.ylim((-1.2, 1.2))
plt.xlabel("Anomaly Score")
plt.title("Predicted Anomaly Score for the Test Set")
plt.show()

Which creates a plot like this We can see that global outliers (zeros) and local outliers (strangely written digits) receive high anomaly scores.

Deployment

docker build -t deep-adots .
docker run -ti deep-adots /bin/bash -c "python3.6 /repo/main.py"

Authors/Contributors

Team:

Supervisors:

Credits

Base implementation for DAGMM
Base implementation for Donut
Base implementation for Recurrent EBM
Downloader for real-world datasets

snowisland1/DeepADoTS