1. 问题的定义：Problem Formulation

image.png image.png image.png

2. 应用

image.png

3. 异常检测可否看作二值分类？

image.png

一是无法穷举所有异常的情况，也就无法知道异常的分布，因为异常的情况变化太大；
二是很难收集到所有的异常示例

4. 异常检测的类别:有无标签，是否数据污染

image.png

4.1 With Label

image.png image.png image.png

如何估计信心分数

image.png

深度学习方法来确定分类和信心分数

image.png image.png

评估标注

image.png image.png

4.2 Without Label

image.png image.png image.png image.png image.png image.png image.png
image.png
image.png

参考文献

========

Open Set Recognition
论文阅读-Open set Recognition

Trying to find task :

https://medium.com/alex-attia-blog/the-simpsons-character-recognition-using-keras-d8e1796eae36

(simple is easy)

https://medium.com/@tyreeostevenson/teaching-a-computer-to-classify-anime-8c77bc89b881

https://github.com/mitmul/chainer-handson/blob/master/animeface-character/classify_characters.ipynb

Pokemeon v.s. difimont

========================

Tool:

https://github.com/yzhao062/anomaly-detection-resources

https://scikit-learn.org/stable/modules/outlier_detection.html

========================

There is an overview paper open-set detector

=====================

https://www.youtube.com/watch?v=12Xq9OLdQwQ (learn some theory here)

https://www.youtube.com/watch?v=5vrY4RbeWkM (too simple)

http://cucis.ece.northwestern.edu/projects/DMS/publications/AnomalyDetection.pdf (old 2009)

Overview: https://murphymind.blogspot.com/2017/07/anomaly.detection.html

NG‘s course: https://www.youtube.com/watch?v=086OcT-5DYI

https://www.quora.com/Do-generative-adversarial-networks-function-for-outlier-detection

https://disp.cc/b/163-b9My

以後看到奇怪的事，就要到派出所報案喔

https://zhuanlan.zhihu.com/p/42267652

Time series data

https://github.com/numenta/NAB

Lots of data set?

http://odds.cs.stonybrook.edu/

https://github.com/alexattia/SimpsonRecognition
https://www.kaggle.com/alexattia/the-simpsons-characters-dataset/kernels

Assumption: most of them are normal player

https://www.reddit.com/r/twitchplayspokemon/comments/49ki48/top_ten_tpp_trolls_of_all_time/

https://jgeekstudies.org/2016/10/08/the-effect-of-trolls-on-twitch-plays-pokemon/

https://www.wired.com/2014/02/twitch-plays-pokemon/

https://kotaku.com/how-people-are-actually-making-progress-in-twitch-play-1525261786

https://kotaku.com/not-everyone-playing-twitch-plays-pokemon-appears-to-1530921548

https://www.quora.com/Do-all-the-trolls-on-Twitch-Plays-Pokemon-make-it-funnier

========

Do I have to mention the native approach like clustering or KNN??????

====================

There is an overview paper open-set detector

=====================
Adversarially Learned One-Class Classifier for Novelty Detection

[1] ](https://www.quora.com/Do-generative-adversarial-networks-function-for-outlier-detection)Artificially intelligent painters invent new styles of art

[2] Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery

[3] http://openaccess.thecvf.com/con...

[4] Shehroz Khan's answer to Could I use GANs to generate negative samples for one class classification?

[5] Satellite Image Forgery Detection and Localization Using GAN and One-Class Classifier