/awesome-anomaly-detection

A complete list of papers on anomaly detection.

Awesome Anomaly Detection[![Awesome]

A list of Papers on anomaly detection. You are welcome to open an issue and pull your requests if you think any paper that is important but not are inclueded in this repo.

Machine Learning Method

This paper is really hard to follow yet some idea are good.

A locally density based method.

Deep Learning Method

Likelihood Generative Methods

Most Auto-encoder methods use either reconstruction error or negative log-likelihood, this is novel. This has been used by Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications in Application section.

Rejected by ICLR2019. But i personally think this is a good paper.

GAN based

RNN

Multivariate Guassian distribution based.

Hypersphereical Learning

Inherit the idea of Anomaly Detection in Dynamic Networks using Multi-view Time-Series Hypersphere Learning, using an extra layer to separate anomaly instances among normal instance.

A well written paper, sold a excellent story, but this paper didn't cited Anomaly Detection in Dynamic Networks using Multi-view Time-Series Hypersphere Learning, which inherited the same idea but published earlier than Deep One-Class Classification. I doubt the level of nnovelty in this paper though it was presented as oral paper in ICML 2018. Additionally, i have read carefully to the codes that were made public by author, the implementation on the loss function of SVDD are different from what they claimed in paper.

One-Class Classification

Hybrid

Incorporate mixture gaussian by energy function.

PCA

Correlation

Ranking

Survey

I read this survey only because this is the sole survey in One Class Classification.

FeedBack

Modifications are made on the basis of Isolation Forest.

An improved version of "Incorporating Feedback into Tree-based Anomaly Detection"

Anomaly Detection Application

KPI

Log

This is the first paper apply deep learning method to detect anomaly in log data. However, there are some obvious flaws in the experiment(i.e. The number of log templates are small, yet the top 10 predicted templates are used as true predications. Additionally, the author refuesed to share more details about there model.)