Authors:

• Saachi Jain - [email protected] - Stanford University
• David Hallac - [email protected] - Stanford University
• Rok Sosic - [email protected] - Stanford University
• Jure Leskovec - [email protected] - Stanford University

Conference:

• KDD 2019, Workshop on Mining and Learning from Time Series

Keywords:

• Noisy motif discovery
• Temporal clustering
• Multivariate time series

Tasks:

• Cluster analysis
• State definition
• Motif discovery

Code:

• https://github.com/snap-stanford/masa.git

Datasets: [TO BE EDITTED]

1. Subjects cycling on an exercise bike, **Daily and Sports Activities Data Set **
2. Aircrafts, commercial & classified
3. Automobiles, commercial & classified

Reviewer:

• Ding Quan

Paper

Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network KDD'19

Code

https://github.com/smallcowbaby/OmniAnomaly

Author

Task

Anomaly Detection for multivariable time series

Dataset

3 public datasets SMAP (Soil Moisture Active Passive satellite), MSL (Mars Science Laboratory rover) and SMD (Server Machine Dataset)

Reviewer

Bin Zhou

Stochastic Recurrent Neural Network

GRU

GRU cell to generate hidden deterministic variable capturing long-term complex temporal information in x-space

Stochastic variable connection

In qnet, model temporal dependence by concatenation.
In pnet, model temporal dependence by Linear Gaussian State Space Model

Plannar Normalizing Flow

Capture complicated data patterns of posterior p(z|x) instead of simple gaussian assumption.

Automatic Threshold Selection

Set the anomaly threshold following the principle of Extreme Value Theory (EVT).
Fit the tail portion of a probability distribution by a generalized Pareto distribution (GPD)
with parameters.

Conference

 WWW 2019

Author:

   Si Zhang
   Hanghang Tong
   Ross Maciejewski
   Tina Eliassi-Rad

Reviewer:

   高浩

Paper

Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications WWW'18

Code

https://github.com/haowen-xu/donut

Author

Task

Unsupervised Anomaly Detection for time series

Dataset

3 business KPI time series (private)

Reviewer

Bin Zhou

Reconstruction by VAE framework

• Modified ELBO
  
  a_w indicate whether x_w is abnormal (0-normal / 1-abnormal)
  β is the shrink factor defined as mean of a_w

• Missing Data Injection
  Randomly set λ ratio of normal points to be zero, as if they are missing points, thus the
  effect of M-ELBO is amplified. (this trick is similar to dropout)

• MCMC imputation
  Since the missing data will introduce the biases when encode the input x, the authors adopt the MCMC-based missing data imputation technique with the trained VAE. With M iterations, we replace missing data with reconstructed data, then use it as the final input x.
  

• Reconstruction probability as anomaly score
  Unlike reconstruction error in autoencoder. The authors use reconstruction probability as anomaly score, defined as:
  

And use Monte Carlo integration to calculate the Expectation

Summary

• Provide the theoretical explanation for VAE-based anomaly detection.
• Modified ELBO and MCMC imputation can significantly improve performance.

Paper

Outlier Detection for Time Series with Recurrent Autoencoder Ensemble IJCAI'19

Code

https://github.com/tungk/OED

Authors

• Tung Kieu
• Bin Yang
• Chenjuan Guo
• Christian S. Jensen

Task

Outlier Detection for Time Series

Datasets

• Numenta Anomaly Benchmark (NAB)
• Electrocardiography (ECG)

RNN ensemble framework

Consider Recurrent Skip Connection Networks (RSCNs) (ref Wang and Tian, 2016).
Set different L for skip connection and the function:

For ensemble framework, we need different structures. So we random remove some connection. Specifically, we introduce a sparseness weight vector , and make sure each at least have one element equal to 1.
The total sparsely-connected RNNs (S-RNNs) can be expressed as

And the figure shows the final structure

Independent Framework

Set different w_t for each autoencoder, and train the autoencoders independently.

Shared Framework

Add a shared layer and concatenate all encoders' codes and use it as init hidden state for all decoder.

Questions

Does anyone know more about different skip connection strategies in RNN?

周厚全: GAN, WGAN

刘财政:
Efficient gan-based anomaly detection
DOPING: Generative Data Augmentation for Unsupervised Anomaly Detection with GAN

周斌：A Survey on GANs for Anomaly Detection
Ganomaly: Semi-supervised anomaly detection via adversarial training
MAD-GAN: Multivariate Anomaly Detection for Time Series Data with Generative Adversarial Networks

张嘉宝
Variational Graph Auto-Encoders, Thomas N. Kipf, Max Welling
Graph Convolutional Matrix Completion

曾四为

报告是注明：会议、日期、google引用次数、作者