If you want an overall understanding of what I was doing, please refer to the PPT

"IML(presentation version).pptx" file. (You will only want to see this)

My project for CIP(Thanks Cathay LAB department for giving me the opportunity! I'm very grateful.)

Folder layout:

1. Data Shapley with image : Trying to use Data Shapley to interpret a neural network
   Data : cats and dogs image

2. InterpreML with relational data : Using Microsoft Interpret to interpret various algorithms
   Data : Bank Marketing dataset(Portugesse)

3. Interpretable Machine Learning : Just a folder storing all the cores that my works use,including Interpret and DataShapley
   and some random tests mainly for debugging

Papers I've read just for this project:

LIME

https://arxiv.org/pdf/1602.04938.pdf
My comment: A very innovative framework for interpreting a model via linear approximate

SHAP

https://arxiv.org/abs/1705.07874
My comment : Extraordinary theory to prove that Shapley Value(Game Theory) is the best method of attributing feature contributions.

Data Shapley

https://arxiv.org/pdf/1904.02868.pdf

Very recent(2019/5) paperwork
My comment:
Another innovative way to implement Shapley Value, SHAP uses Features as players(in game theory),while Data Shapley uses data instances as players.
Data Shapley我認為真正進入到狹義的Global Interpretation，即可以得知＂預測模型＂(v.s.解釋模型)預測出來的結果大概是長什麼樣子
Local interpretation我個人理解為＂得知預測模型怎麼預測(i.e.使用哪些features及相關的權重，且很好解釋)

經過資料測試，發現只要跑稍為複雜的模型(bi-layer neural net work with 100 hidden units)，跑個1000筆資料就足以耗上好幾天的時間(on AWS,xlarge instance)
Cons : 因為time complexity實在太高，只適合用在較珍貴的資料集上，來算出各資料點的價值

Microsoft Interpret

An overall interpret Opensource package by Microsoft(has SHAP/LIME/EBM) , a very nice visualization tool for interpretation
Internal Engine : Explainable Boosting Machine(EBM) (GA2M Paper) : http://www.cs.cornell.edu/~yinlou/papers/lou-kdd13.pdf
Github: https://github.com/microsoft/interpret

My comment:
屬於統計學習一派，Microsoft所開發出來（與想推銷的）內部引擎EBM為一個可解釋的＂預測模型＂，使用統計學習方法，以GAM為Baseline Model，將GAM回歸出來的殘差，再做一次交叉項(Pairwise Interaction effect)的回歸，期望把更多的殘差以交互作用來納入整體解釋能力。
但個人感覺光是GAM可能就很難解釋了(Smoother function，你所train出來的個別函數是什麼等等)，加上GA2M又更難解釋，見仁見智。

一些pdf檔案為我看過之後加上一些自己的筆記，有需要拿就自取吧

Last :
a very good book for Interpretable Machine Learning(online,free), I learned very very much from this: https://christophm.github.io/interpretable-ml-book/

Version Control

8/19/2019

Upload InterpretML with relational data

Using bank marketing datasets from https://archive.ics.uci.edu/ml/datasets/Bank+Marketing
Applies Microsoft InterpretML on it

Upload Data Shapley with image (Most time consuming here)

Using datasets from https://github.com/Pranaw99/Image_Classification_CNN
Applies Data Shapley on cat/dog images and compute the Data Shapley value
這邊還沒完整測試過，跑一次Data Shapley需要大量時間，非常難debug
所以正確性還有待觀察，持續補強中
使用前且確保清除該路徑所有temp、.ipynb_checkpoints與__pycache__資料夾(執行DShap會產生)，不然發生一些奇妙的錯誤(這邊我花了好幾天百思不得其解之後才知道應該是快取問題)

8/26/2019

Upload Data Shapley image with CNN (version3)

using 250 data points
Too time consuming, on AWS xlarge instance, it has to run at least 5 days or more
I didnt finished training this model,but it should work. Feel free to use if you want

8/30/2019

Final upload, a small change to README, and add final powerpoint file