School of Computer Science and Electronics Engineering, University of Essex

IInroduction to Explainable AI in Machine Learning

CE880: An Approachable Introduction to Data Science

Mohsin Ali
14 March 2023

1
About Myself

I Name: Mohisn Ali

I Position: PhD Scholar
I Research interest: Image Processing, Deep Learning and Machine Vision
I Contact: ma22159@essex.ac.uk

2
Why Explainable AI?

Machine Learning models are often black boxes that make their interpretation difficult,
even by the creator of that model. In order to understand why a particular model is
returning some prediction, it is important to understand the main features that affect
the model’s output.

0
Topbots

3
Clever Hans

In 1907 there was a horse (Clever Hans) that was claimed to have performed
arithmetic and other intellectual tasks

0
wikipedia

4
Does the highest accuracy mean the best model?

Examples:

I Covid Detection
I Cancer Detection
I Fraud Detection

5
Real-life Examples | Amazon

By 2015, Amazon realized its new system was not rating candidates for software
developer jobs and other technical posts in a gender-neutral way.

0
Reuters

6
Real-life Examples | IBM, Microsoft

Face recognition algorithms boast high classification accuracy (over 90%), but these
outcomes are not universal. A growing body of research exposes divergent error rates
across demographic groups, with the poorest accuracy consistently found in subjects
who are female, Black, and 18-30 years old.
0
Harvard University, BBC

7
Further Explanation

Coded Bias lists actual examples of algorithms already making decisions about your
credit, your health, your housing, your college or job applications, even your access to
possibility. The system holds in its hands your hope of a better life – and there’s no
appeal if the computer says no.

0
Netflix, MIT, Cnet

8
What is Explainable AI (XAI)

Explainable artificial intelligence (XAI) is a set of processes and methods that allows
human users to understand the results generated by machine learning algorithms.
Which helps to:

I Build trust
I Meet the regulatory standards
I Acquire fairness

9
Explainable AI (XAI) Process

10
Types of Explainable AI methods

11
SHAP (SHapley Additive exPlanations)

SHAP (SHapley Additive exPlanations) is a game theoretic approach to explain the

output of any machine learning model. Using the classic Shapley values from game
theory.

0
https://github.com/slundberg/shap

12
SHAP Explanation

0
https://github.com/slundberg/shap

13
Game Theory

In the 1950s John Nash proposed the game theory.

A game is any social interaction between multiple people where each person’s reward
is affected by other people’s decisions.

14
Type of Game Theory

There are 2 types of game theory:

I Non-Cooperative Game Theory

I Cooperative Game Theory

15
Non-Cooperative Game Theory | Prisoners Dilemma

16
Non-Cooperative Game Theory

Non-Cooperative Game Theory state that a player has found Nash Equilibrium when
they make that choice that leaves them better off no matter what their opponents
decide to do

17
Cooperative Game Theory

I Cooperative Games are games where all players agree to work to archive a
common goal
I Cooperative Game Theory tries to find how much each player has contributed to
the game and how much reward each player should receive
I Similar, to Non-Cooperative Game Theory where there is Nash Equilibrium
Cooperative Game Theory has shapely values

18
Shapley Values

I The contribution of each player is determined by what they lost or gained after
removing them from the game (Marginal Contribution)
I Interchangeable players have equal values
I Dummy players should have zero values
I If a game has multiple parts reward should be calculated for each part separately

19
Shapley Values | Example

20
Shapley Values | Example

21
Shapley Values | Example

Shapely values state that you should get 15$ and your friend should get 25$ from 40$

22
Titanic

On April 15, 1912, during her maiden voyage, the widely considered “unsinkable” RMS
Titanic sank after colliding with an iceberg. Unfortunately, there weren’t enough
lifeboats for everyone onboard, resulting in the death of 1502 out of 2224 passengers
and crew.
While there was some element of luck involved in surviving, it seems some groups of
people were more likely to survive than others.

0
Wikipedia, Kaggle

23
Titanic Dataset

0
Kaggle

24
Install SHAP

Shap can be installed from either PyPI:

0
https://shap-lrjball.readthedocs.io/

25
Downloading the dataset

Download the dataset from https://www.kaggle.com/competitions/titanic/data

0
Kaggle

26
Load Dataset

Load the data as pandas dataframe

Compress the train and test data into 1 list

27
Change the String data to Numerical data

28
Data Information

29
Fill null values

Fill null values with the median of the column

30
Verify Data

Verify the data for null values

31
Continuous data to categorical data

We will convert the continuous data to categorical data for a better understanding

32
Continuous data to categorical data

33
String data to Numerical

Convert Embarked port to Numeric data

Convert Title to Numeric data

34
Remove unnecessary features

35
Fit Machine Learning model

36
Get SHAP values

37
Plot SHAP | Summary Plot | Bar

38
Plot SHAP | Summary Plot | Beeswarm

39
Plot SHAP | Individual Instance | Bar

40
Plot SHAP | Individual Instance | Waterfall

41
Returning back to ’Titanic’

Why Jack did not survive?

42
CE880: An Approachable Introduction to Data Science

43