AAISI AI Colloquium 30/3/2021: Bias in AI systems

AI Colloquium by AAISI (online), 30.3.2021
Bias in AI systems: A multi-step approach
Eirini Ntoutsi
Free University Berlin

Outline
 Introduction
 Dealing with bias in data-driven AI systems
 Understanding bias
 Mitigating bias
 Accounting for bias
 Case: bias-mitigation with sequential ensemble learners (boosting)
 Wrapping up
3
Eirini Ntoutsi Bias in AI systems: A multi-step approach

Successful applications
4
Recommendations Navigation
Severe weather alerts
Automation

Questionable uses/ failures
5
Google flu trends failure Microsoft’s bot Tay taken offline
after racist tweets
IBM’s Watson for Oncology
cancelled
Facial recognition works better
for white males

Why AI-projects might fail?
 Back to basics: How machines learn
 Machine Learning gives computers the ability to learn without being
explicitly programmed (Arthur Samuel, 1959)
 We don’t codify the solution, we don’t even know it!
 DATA & the learning algorithms are the keys
6
Algorithms
Models
Models
Data

Mind the (hidden) assumptions
 Assumptions include: stationarity, independent & identically distributed
(iid) data, balanced class representation, ...
 In this talk, I will focus on the assumption/myth of algorithmic objectivity
1. The common misconception that humans are subjective, but data and
algorithms not and therefore they cannot discriminate.
7

Reality check: Can algorithms discriminate?
 Bloomberg analysts compared Amazon same-day delivery areas with U.S.
Census Bureau data
 They found that in 6 major same-day delivery cities, the service area
excludes predominantly black ZIP codes to varying degrees.
 Shouldn’t this service be based on customer’s spend rather than race?
 Amazon claimed that race was not used in their models.
8
Source: https://www.bloomberg.com/graphics/2016-amazon-same-day/

Reality check cont’: Can algorithms discriminate?
 There have been already plenty of cases of algorithmic discrimination
 State of the art visions systems (used e.g. in autonomous driving) recognize
better white males than black women (racial and gender bias)
 Google’s AdFisher tool for serving personalized ads was found to serve
significantly fewer ads for high paid jobs to women than men (gender-bias)
 COMPAS tool (US) for predicting a defendant’s risk of committing another
crime predicted higher risks of recidivism for black defendants (and lower for
white defendants) than their actual risk (racial-bias)
9

Dont blame (only) the AI
 “Bias is as old as human civilization” and “it is human nature for members
of the dominant majority to be oblivious to the experiences of other
groups”
 Human bias: a prejudice in favour of or against one thing, person, or group
compared with another usually in a way that’s considered to be unfair.
 Bias triggers (protected attributes): ethnicity, race, age, gender, religion, sexual
orientation …
 Algorithmic bias: the inclination or prejudice of a decision made by an AI
system which is for or against one person or group, especially in a way
considered to be unfair.
10

Every bias is not necessarily a bad bias 1/2
 Inductive bias ”refers to a set of (explicit or implicit) assumptions made by
a learning algorithm in order to perform induction, that is, to generalize a
finite set of observation (training data) into a general model of the
domain. Without a bias of that kind, induction would not be possible,
since the observations can normally be generalized in many ways.”
(Hüllermeier, Fober & Mernberger, 2013)
 Bias-free learning is futile: A learner that makes no a priori assumptions
regarding the identity of the target concept has no rational basis for
classifying any unseen instances.
11
Model
Future unknown instances

Every bias is not necessarily a bad bias 2/2
 Some biases are positive and helpful, e.g., making healthy eating choices
 Some biases help us to become more efficient
 E.g., start work early if you are a morning person
 We refer here to bias that might cause discrimination and unfair actions
to an individual or group on the basis of protected attributes like race or
gender.
 Bias and gender are examples of bias triggers, but not the only ones
 Combined triggers as well, e.g., black females in IT jobs
 Bias and discrimination depend on context
 Women in tech
 Males in ballet dancing
12

Outline
 Introduction
 Mitigating bias
 Wrapping up
13

The fairness-aware machine learning domain
 A young, fast evolving, multi-disciplinary research field
 Bias/fairness/discrimination/… have been studied for long in philosophy, social
sciences, law, …
 Existing approaches can be divided into three categories
 How bias is created in the society and enters our sociotechnical systems, is
manifested in the data used by AI algorithms, and can be formalized.
 Mitigating bias
 Approaches that tackle bias in different stages of AI-decision making.
 Approaches that account for bias proactively or retroactively.
14
E. Ntoutsi, P. Fafalios, U. Gadiraju, V. Iosifidis, W. Nejdl, M.-E. Vidal, S. Ruggieri, F. Turini, S. Papadopoulos, E. Krasanakis, I. Kompatsiaris, K. Kinder-
Kurlanda, C. Wagner, F. Karimi, M. Fernandez, H. Alani, B. Berendt, T. Kruegel, C. Heinze, K. Broelemann, G. Kasneci, T. Tiropanis, S. Staab"Bias in
data-driven artificial intelligence systems—An introductory survey", WIREs Data Mining and Knowledge Discovery, 2020.

Outline
 Introduction
 Mitigating bias
 Wrapping up
16

Understanding bias: Sociotechnical causes of bias
 AI-systems rely on data generated by humans (UGC) or collected via
systems created by humans.
 As a result human biases
 enter AI systems
 E.g., bias in word-embeddings (Bolukbasi et al, 2016)
 might be amplified by complex sociotechnical systems such as the Web and,
 new types of biases might be created
17

Understanding bias: How is bias manifested in data?
 Protected attributes and proxies
 E.g., neighborhoods in U.S. cities are highly correlated with race
 Representativeness of data
 E.g., underrepresentation of women and people of color in IT developer
communities and image datasets
 E.g., overrepresentation of black people in drug-related arrests
 Depends on data modalities
18
https://incitrio.com/top-3-lessons-learned-from-
the-top-12-marketing-campaigns-ever/
https://ellengau.medium.com/emily-in-
paris-asian-women-i-know-arent-like-
mindy-chen-6228e63da333

Typical (batch) fairness-aware learning setup
 Input: D = training dataset drawn from a joint distribution P(F,S,y)
 F: set of non-protected attributes
 S: (typically: binary, single) protected attribute
 s (s ̄): protected (non-protected) group
 y = (typically: binary) class attribute {+,-} (+ for accepted, - for rejected)
 Goal of fairness-aware classification: Learn a mapping from f(F, S) → y
 achieves good predictive performance
 eliminates discrimination
19
F1 F2 S y
User1 f11 f12 s +
User2 f21 -
User3 f31 f23 s +
… … … … …
Usern fn1 +
We know how to measure this
According to some fairness measure

Measuring (un)fairness: some measures
 Statistical parity: If subjects in both protected and unprotected groups
should have equal probability of being assigned to the positive class
𝑃 ො
𝑦 = + 𝑆 = 𝑠 = 𝑃 ො
𝑦 = + 𝑆 = ҧ
𝑠
 Equal opportunity: There should be no difference in model’s prediction
errors regarding the positive class
𝑃 ො
𝑦 ≠ 𝑦 𝑆 = 𝑠+ = 𝑃 ො
𝑦 ≠ 𝑦 𝑆 = ҧ
𝑠+
 Disparate Mistreatment: There should be no difference in model’s
prediction errors between protected and non-protected groups for both
classes
𝛿FNR = 𝑃 ො
𝑦 ≠ 𝑦 𝑆 = 𝑠+ − 𝑃 ො
𝑦 ≠ 𝑦 𝑆 = ҧ
𝑠+
𝛿FPR = 𝑃 ො
𝑦 ≠ 𝑦 𝑆 = 𝑠− − 𝑃 ො
𝑦 ≠ 𝑦 𝑆 = ҧ
𝑠−
Disparate Mistreatment = 𝛿𝐹𝑁𝑅 + 𝛿𝐹𝑃𝑅
22
F1 F2 S y ෝ
𝒚
User1 f11 f12 s + -
User2 f21 - +
User3 f31 f23 s + -
… … … … … …
Usern fn1 + +

Outline
 Introduction
 Mitigating bias
 Wrapping up
24

Mitigating bias
 Goal: tackling bias in different stages of AI-decision making
26
Algorithms
Models
Models
Data
Applications
Hiring
Banking
Healthcare
Education
Autonomous
driving
…
Pre-processing
approaches
In-processing
approaches
Post-processing
approaches

Mitigating bias: pre-processing approaches
 Intuition: making the data more fair will result in a less unfair model
 Idea: balance the protected and non-protected groups in the dataset
 Design principle: minimal data interventions (to retain data utility for the
learning task)
 Different techniques:
 Instance class modification (massaging), (Kamiran & Calders, 2009),(Luong,
Ruggieri, & Turini, 2011)
 Instance selection (sampling), (Kamiran & Calders, 2010) (Kamiran & Calders,
2012)
 Instance weighting, (Calders, Kamiran, & Pechenizkiy, 2009)
 Synthetic instance generation (Iosifidis & Ntoutsi, 2018)
 …
27

Mitigating bias: pre-processing approaches: Massaging
 Change the class label of carefully selected instances (Kamiran & Calders, 2009).
 The selection is based on a ranker which ranks the individuals by their probability to
receive the favorable outcome.
 The number of massaged instances depends on the fairness measure (group fairness)
28
Image credit Vasileios Iosifidis

Mitigating bias
30
Algorithms
Models
Models
Data
Applications
Hiring
Banking
Healthcare
Education
Autonomous
driving
…
Pre-processing
approaches
In-processing
approaches
Post-processing
approaches

Mitigating bias: in-processing approaches
 Intuition: working directly with the algorithm allows for better control
 Idea: explicitly incorporate the model’s discrimination behavior in the
objective function
 Design principle: “balancing” predictive- and fairness-performance
 Regularization (Kamiran, Calders & Pechenizkiy, 2010),(Kamishima, Akaho,
Asoh & Sakuma, 2012), (Dwork, Hardt, Pitassi, Reingold & Zemel, 2012) (Zhang
& Ntoutsi, 2019)
 Constraints (Zafar, Valera, Gomez-Rodriguez & Gummadi, 2017)
 Training on latent target labels (Krasanakis, Xioufis, Papadopoulos &
Kompatsiaris, 2018)
 In-training altering of data distribution (Iosifidis & Ntoutsi, 2019)
 …
31

Mitigating bias: in-processing approaches: change the objective
function 1/2
 An example with decision tree (DTs) classifiers
 Traditional DTs use pureness of the data (in terms of class-labels) to decide on
which attribute to use for splitting
 Which attribute to choose for splitting? A1 or A2?
 The goal is to select the attribute that is most useful for classifying examples.
 helps us be more certain about the class after the split
 we would like the resulting partitioning to be as pure as possible
 Pureness evaluated via entropy - A partition is pure if all its instances belong to the same class.
 But traditional measures (Information Gain, Gini Index, …) ignore fairness
32

Mitigating bias: in-processing approaches: change the objective
function 2/2
 We introduce the fairness gain of an attribute (FG)
 Disc(D) corresponds to statistical parity (group fairness)
 We introduce the joint criterion, fair information gain (FIG) that evaluates
the suitability of a candidate splitting attribute A in terms of both
predictive performance and fairness.
33
D
D1
D2
W. Zhang, E. Ntoutsi, “An Adaptive Fairness-aware Decision Tree Classifier", IJCAI 2019.

Mitigating bias
34
Algorithms
Models
Models
Data
Applications
Hiring
Banking
Healthcare
Education
Autonomous
driving
…
Pre-processing
approaches
In-processing
approaches
Post-processing
approaches

Mitigating bias: post-processing approaches
 Intuition: start with predictive performance
 Idea: first optimize the model for predictive performance and then tune
for fairness
 Design principle: minimal interventions (to retain model predictive
performance)
 Correct the confidence scores (Pedreschi, Ruggieri, & Turini, 2009), (Calders &
Verwer, 2010)
 Correct the class labels (Kamiran et al., 2010)
 Change the decision boundary (Kamiran, Mansha, Karim, & Zhang, 2018), (Hardt,
Price, & Srebro, 2016)
 Wrap a fair classifier on top of a black-box learner (Agarwal, Beygelzimer, Dudík,
Langford, & Wallach, 2018)
 …
35

Mitigating bias: pοst-processing approaches: shift the decision
boundary
 An example of decision boundary shift
36
V. Iosifidis, H.T. Thi Ngoc, E. Ntoutsi, “Fairness-enhancing interventions in stream classification", DEXA 2019.

Outline
 Introduction
 Mitigating bias
 Wrapping up
37

Accounting for bias
 Algorithmic accountability refers to the assignment of responsibility for
how an algorithm is created and its impact on society (Kaplan et al, 2019).
 Many facets of accountability for AI-driven algorithms and different
approaches
 Proactive approaches:
 bias-aware data collection, e.g., for Web data, crowd-sourcing
 Bias-description and modeling, e.g., via ontologies
 ...
 Retroactive approaches:
 Explaining AI decisions in order to understand whether decisions are biased
 What is an explanation? Explanations w.r.t. legal/ethical grounds?
 Using explanations for fairness-aware corrections (inspired by Schramowski et al, 2020)
38

Outline
 Introduction
 Mitigating bias
 Wrapping up
40

Fairness with sequential learners (boosting)
 Sequential ensemble methods generate base learners in a sequence
 The sequential generation of base learners promotes the dependence between
the base learners.
 Each learner learns from the mistakes of the previous predictor
 The weak learners are combined to build a strong learner
 Popular examples: Adaptive Boosting (AdaBoost), Extreme Gradient Boosting
(XGBoost).
 Our base model is AdaBoost (Freund and Schapire, 1995), a sequential ensemble
method that in each round, re-weights the training data to focus on misclassified
instances.
41
Round 1: Weak learner h1 Round 2: Weak learner h2 Round 3: Weak learner h3 Final strong learner H()

Intuition behind using boosting for fairness
 It is easier to make “fairness-related interventions” in simpler models
rather than complex ones
 We can use the whole sequence of learners for the interventions instead
of the current one
44

Limitations of related work
 Existing works evaluate predictive performance in terms of the overall
classification error rate (ER), e.g., [Calders et al’09, Calmon et al’17, Fish et
al’16, Hardt et al’16, Krasanakis et al’18, Zafar et al’17]
 In case of class-imbalance, ER is misleading
 Most of the datasets however suffer from imbalance
 Moreover, Dis.Mis. is “oblivious” to the class imbalance problem
47

From Adaboost to AdaFair
 We tailor AdaBoost to fairness
 We introduce the notion of cumulative fairness that assesses the fairness of
the model up to the current boosting round (partial ensemble).
 We directly incorporate fairness in the instance weighting process
(traditionally focusing on classification performance).
 We optimize the number of weak learners in the final ensemble based on
balanced error rate thus directly considering class imbalance in the best model
selection.
48
𝐸𝑅 = 1 −
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝐹𝑁 + 𝑇𝑁 + 𝐹𝑃
V. Iosifidis, E. Ntoutsi, “AdaFair: Cumulative Fairness Adaptive Boosting", ACM CIKM 2019.

AdaFair: Cumulative boosting fairness
 Let j: 1−T be the current boosting round, T is user defined
 Let be the partial ensemble, up to current round j.
 The cumulative fairness of the ensemble up to round j, is defined based
on the parity in the predictions of the partial ensemble between
protected and non-protected groups
 ``Forcing’’ the model to consider ``historical’’ fairness over all previous
rounds instead of just focusing on current round hj() results in better
classifier performance and model convergence.
50

AdaFair: fairness-aware weighting of instances
 Vanilla AdaBoost already boosts misclassified instances for the next round
 Our weighting explicitly targets fairness by extra boosting discriminated
groups for the next round
 The data distribution at boosting round j+1 is updated as follows
 The fairness-related cost ui of instances xi ϵ D which belong to a group
that is discriminated is defined as follows:
51

AdaFair: optimizing the number of weak learners
 Typically, the number of boosting rounds/ weak learners T is user-defined
 We propose to select the optimal subsequence of learners 1 … θ, θ ≤ T
that minimizes the balanced error rate (BER)
 In particular, we consider both ER and BER in the objective function
𝑎𝑟𝑔𝑚𝑖𝑛𝜃 𝑐 ∗ 𝐵𝐸𝑅𝜃 + 1 − 𝑐 𝐸𝑅𝜃 + 𝑀𝑖𝑠. 𝐷𝑖𝑠.
 The result of this optimization if a final ensemble model with Eq.Odds
fairness
53

Experimental evaluation
 Datasets of varying imbalance
 Baselines
 AdaBoost [Sch99]: vanilla AdaBoost
 SMOTEBoost [CLHB03]: AdaBoost with SMOTE for imbalanced data.
 Krasanakis et al. [KXPK18]: Boosting method which minimizes Dis.Mis. by approximating
the underlying distribution of hidden correct labels.
 Zafar et al.[ZVGRG17]: Training logistic regression model with convex-concave
constraints to minimize Dis.Mis.
 AdaFair NoCumul: Variation of AdaFair that computes the fairness weights based on
individual weak learners.
54

Experiments: Predictive and fairness performance
 Adult census income (ratio 1+:3-)  Bank dataset (ratio 1+:8-)
Eirini Ntoutsi Bias in AI systems: A multi-step approach 55
Larger values are better, for Dis.Mis. lower values are better
 Our method achieves high balanced accuracy and low discrimination (Dis.Mis.) while maintaining high
TPRs and TNRs for both groups.
 The methods of Zafar et al and Krasanakis et al, eliminate discrimination by rejecting more positive
instances (lowering TPRs).

Cumulative vs non-cumulative fairness
 Cumulative vs non-cumulative fairness impact on model performance
 Cumulative notion of fairness performs better
 The cumulative model (AdaFair) is more stable than its non-cumulative
counterpart (standard deviation is higher)
56
Please note: Eq.Odds Dis.Mis.

Outline
 Introduction
 Mitigating bias
 Wrapping up
57

Wrapping-up, ongoing work and future directions
 In this talk I focused on the myth of algorithmic objectivity and
 the reality of algorithmic bias and discrimination and how algorithms can pick biases
existing in the input data and further reinforce them
 A large body of research already exists but
 focuses mainly on fully-supervised batched learning with single-protected (and typically
binary) attributes with binary classes
 Moving from batch learning to online learning
 targets bias in some step of the analysis-pipeline, but biases/errors might be propagated
and even amplified (unified approached are needed)
 Moving from isolated approaches (pre-, in- or post-) to combined approaches
58
V. Iosifidis, E. Ntoutsi, “FABBOO - Online Fairness-aware Learning under Class Imbalance", DS 2020.
T. Hu, V. Iosifidis, W. Liao, H. Zang, M. Yang, E. Ntoutsi,B. Rosenhahn, "FairNN - Conjoint Learning of Fair Representations for Fair
Decisions”, DS 2020.

Wrapping-up, ongoing work and future directions
 Moving from single-protected attribute fairness-aware learning to multi-
fairness
 Existing legal studies define multi-fairness as compound, intersectional and
overlapping [Makkonen 2002].
 Moving from fully-supervised learning to unsupervised and reinforcement
learning
 Moving from myopic (maximize short-term/immediate performance) solutions
to non-myopic ones (that consider long-term performance) [Zhang et al,2020]
 Actionable approaches (counterfactual generation)
59

Thank you for you attention!
60
Questions?
https://nobias-project.eu/
@NoBIAS_ITN
https://lernmint.org/
Feel free to contact me:
• eirini.ntoutsi@fu-berlin.de
• @entoutsi
• https://www.mi.fu-
berlin.de/en/inf/groups/ag-KIML/index.html
https://www.bias-project.org/

AAISI AI Colloquium 30/3/2021: Bias in AI systems

More Related Content

What's hot

Similar to AAISI AI Colloquium 30/3/2021: Bias in AI systems

More from Eirini Ntoutsi

Recently uploaded

AAISI AI Colloquium 30/3/2021: Bias in AI systems