Keynote: Bias in Search and Recommender Systems

Ricardo Baeza-Yates © | @polarbearby | #TechSEOBoost
#TechSEOBoost | @CatalystSEM
THANK YOU TO THIS YEAR’S SPONSORS
Keynote: Bias in Search and Recommender
Systems
Ricardo Baeza-Yates, NTENT

Ricardo Baeza-Yates
CTO, NTENT
Biases in Search &
Recommender Systems

NTENT ntent.com
Marketing Engineering Operations International Applied Research
ntent.com

Prologue

A Bit of History
Data Volume
Complexity
IR
DB
Two different points
of view for data

Data Understanding
Data
Query
Unstructured Structured
Explicit
Information
Retrieval
(Relational)
Databases
Implicit Recommender Systems
Unknown Data Mining

What is Bias?
• Statistical: significant systematic
deviation from a prior (unknown)
distribution;
• Cultural: interpretations and judgments
phenomena acquired through our life;
• Cognitive: systematic pattern of deviation
from norm or rationality in judgment;

So (Observational) Human Data has Bias
• Gender
• Racial
• Sexual
• Age
• Religious
• Social
• Linguistic
• Geographic
• Political
• Educational
• Economic
• Technological
▪ Gathering process
▪ Sampling process
▪ Validity (e.g. temporal)
▪ Completeness
▪ Noise, spam
Many people extrapolate results of
a sample to the whole population
(e.g., social media analysis)
In addition there is bias when
measuring bias as well as bias
towards measuring it!
Attempt of an unbiased (personal) view on bias in Search & RS
Cultural Biases Statistical Biases Cognitive Biases
Self-selection

Impact in Search and Recommender Systems
• Most web systems are optimized by using implicit user feedback
• However, user data is partly biased to the choices that these systems make
• Clicks can only be done on things that are shown to us
• As those systems are usually based in ML, they learn to reinforce their own biases,
yielding self-fulfilled prophecies and/or sub-optimal solutions
• For example, personalization and filter bubbles for users
• but also echo chambers for (recommender) systems
• Moreover, sometimes these systems compete among themselves, learning also
biases of other systems rather than real user behavior
• Even more, an improvement in one system might be just a degradation in another
system that uses a different (even inversely correlated) optimization function
• For example, user experience vs. monetization

A Non-Technical Question
Algorithm
Biased
Data
Neutral?
Fair?
Same
Bias
Garbage In
Garbage Out

What is being fair?

A Non-Technical Question
Algorithm
Biased
Data
Neutral?
Fair?
Same
Bias
Not
Always!
Debias the input
Tune the algorithm
Debias the output
Bias awareness!

January 2017 ACM US Statement on
Algorithm Transparency and Accountability
1. Awareness
2. Access and redress
3. Accountability
4. Explanation
5. Data Provenance
6. Auditability
7. Validation and Testing
Systems do not need to be perfect,
they just need to be better than us

Biases Everywhere!

Data bias
Biases on Search & RS: Web Case Study
Web Spam

[Baeza-Yates, Castillo & López,
Cybermetrics, 2005]
Number of linked domains
Exports(thousandsofUS$)
Economic Bias in Links

17
[Baeza-Yates & Castillo, WWW2006]
Economic Bias in Links

18
Minimal effortShameCultural Bias in Websites
[Baeza-Yates, Castillo,
Efthimiadis, TOIT 2007]

Language Bias

[Bolukbasi at al, NIPS 2016]
• Word embedding’s in w2vNEWS
Yes, about 60 to 70% at work
although at college is the inverse
Gender Bias in Content
Most journalists are men?

[E. Graells-Garrido et al,.
ACM Hypertext’15]
Systemic bias?
Equal opportunity?
Gender Bias in Content
Wikipedia
Partial
information

Data bias
Activity bias
Bias on Usage
Actions
People
We are all in the long tail!
[Goel et al., WSDM 2010]

Popularity Bias in Recommender Systems
Items
Users
Popular
items
Rest of items
(long tail)
• Take care to recommended items that are
not too popular
• Metrics
• Novelty enhancement
• Problem solved! …really?
𝑛𝑜𝑣 𝑖 = 1 −
# ratings of 𝑖
# users
Items
#interactions
More
novel
Less
novel
𝑎 𝑏
[Vargas & Castells, RecSys 2011]

A Self-fulfilling Prophecy?
Popular items
(short head)
Rest of items
(long tail)
Observed user-item interaction
Unobserved preference
Items
Users
Ratings are missing
not at random (MNAR)

Test data (relevant items)
Training data
Items
Users
Popular items
(short head)
Rest of items
(long tail)
avg P@𝑘 ∼
+
𝑘
Judgments are missing
not at random (MNAR)
Worse yet: user-system
reinforcement loop (more later)
A Self-fulfilling Prophecy?
[Marlin et al., RecSys 2010]
[Steck, RecSys 2010, 2011]
[Fleder & Hossanagar,
Management Sciences 2009]

1.E-06
1.E-05
1.E-04
1.E-03
0 0.2 0.4 0.6 0.8 1
A Problem for IR Evaluation Methodology!
30 TREC collections
Items / documents
#ratings/judgments
(infraction)
[Bellogín, Castells & Cantador, IRJ 2017]
To how many queries
is a document relevant?
25% queries can be
answered with less than
1% of the URLs!
[Baeza-Yates, Boldi,
Chierichetti, WWW 2015]

Get Rid of the Popularity Bias!
• In the rating split
[Bellogín, Castells & Cantador, IRJ 2017]
• In the metrics
• Stratified recall
[Steck, RecSys 2011]
• Importance propensity scoring
[Yang et al., RecSys 2018]
• In the algorithms
[Steck, RecSys 2011]
[Lobato et al., ICML 2014]
[Jannach et al., UMUAI 2015]
[Cañamares & Castells, SIGIR 2018,
best paper award]
Test data (relevant items)
Training data
Items Items
#ratings
Flat test Popularity strata
#ratings
Time
Temporal
split

Activity Bias also Affects Content
[Baeza-Yates & Saez-Trumper, ACM Hypertext 2015]
Most users are passive (i.e., more than 90% are lurkers)
Then, which percentage of active users produce 50% of the content?

October 2015

[Baeza-Yates & Saez-Trumper, ACM Hypertext 2015]
Which percentage of active users produce 50% of the content?
Wisdom of crowds is a partial illusion
Activity Bias: The Wisdom of a Few

Content Never Seen
[Baeza-Yates & Saez-Trumper,
ACM Hypertext 2015]
The Digital Desert

Data bias
Activity bias
Sampling
(size) bias
Algorithmic bias
Search or
Recommender System
Bias on the Web

[E. Graells-Garrido & M. Lalmas, ACM Hypertext’14]
Geographical Bias in Recommender Systems

• If we want to estimate the frequency of queries that appear with probability at
least p with a certain relative error, ∊, we can use the standard binomial error
formula which works well for p near ½ but not for p near 0
• Better is the Agresti-Coull technique (also called Take 2) which gives:
where Z is the inverse of the standard normal distribution,
1 − 𝛼 is the confidence interval and
• If p = 0.1, 1 − 𝛼 is 80% and ∊ is 10%, the standard formula
gives n = 900, while with A-C we get n = 2342.
[Brown, Cai & DasGupta, Statistical Science, 2001]
[Baeza-Yates, SIGIR 2015, Industry track]
Sample Size?

• Standard technique:
• A good sample should cover well all the items distribution, but this
does not work with very skewed distributions.
10
0
10
1
10
2
10
3
Rank
10
0
10
1
10
2
10
3
10
4
10
5
Frequency
50M
10M
100K
1K
10
0
10
1
10
2
10
3
10
4
10
5
10
6
10
7
Rank
10
0
10
1
10
2
10
3
10
4
10
5
Frequency
50M
10M
100K
1K
[Zaragoza et al, CIKM 2010]
Sampling Queries

36
Stratified Sampling Example [Baeza-Yates, SIGIR 2015,
Industry track]

Data bias
Activity bias
Sampling
bias
Interaction bias
(Self) selection bias
Bias in the User Interaction
Search or
Recommender System

Position bias
Ranking bias
Presentation or exposure bias
Social bias
Interaction bias
Bias in the Interaction
Amazon.com

Position bias
Presentation
bias
Social bias
Interaction bias
Ranking bias
Click bias
Scrolling bias
Mouse
movement
bias
Data and algorithmic bias Self-selection bias
Dependencies: A Cascade of Biases!

Ranking Bias in Web Search
[Mediative Study, 2014]

Ranking Bias: Click Bias in Web Search
• Ranking & next page bias
Navigational queries

CTR
(log)
1 11 21 Rank
Learning to Rank with bias
[Joachims et al., WSDM 2017, best paper]
+ many other papers
Fair rankings
[Zehlike et al., CIKM 2017]
Clicks as implicit positive user feedback
Debiasing Search Clicks and Other Biases
[Dupret & Piwowarski, SIGIR 2008]
[Chapelle & Zhang, WWW 2009]
[Dupret & Liao, WSDM 2010, best paper]
Debias the input
Tune the algorithm
Debias the output

Data bias
Activity bias
Sampling
bias
Interaction bias
Second-order bias
Vicious Cycle of Bias
Search or
Recommender System

[Baeza-Yates, Pereira & Ziviani, WWW 2008]
Person
Web content is redundant (> 20%)
Query
Ranking bias in new content
Redundancy grows (35%)
Search results
New page
Second Order Bias in Web Content
[Fortunato, Flammini, Menczer & Vespignani. PNAS 2006]

Data bias
Activity bias
Sampling
bias
Interaction bias
Vicious Cycle of Bias
Search or
Recommender System
Feedback loop bias

Bias due to Personalization
• The effect of self-selection bias
• Avoid the rich get richer and
poor get poorer syndrome
• Avoid the echo chamber by
empowering the tail
Cold start problem solution: Explore & Exploit
Partial solutions:
• Diversity
• Novelty
• Serendipity
• My dark side Wikipedia
[Eli Pariser, Penguin 2011]

Users’ Eco Chambers in Recommender Systems
▪ Filter bubbles
▪ Degenerate feedback loops
(e.g., YouTube autoplay)
[Jiang et al., AAAI 2019]

Eco Chamber of the Recommender System
• Short-term greedy optimization, partial knowledge of the world
• Long-term revenue optimization is not achieved
• Views from new users should balance the exploration for new items
• Disparate impact: unfair ecommerce/information markets
[Baeza-Yates & Ribaudo,
to appear]

Epilogue

Recap
Bias Type Statistical Cultural Cognitive
Algorithmic  ? ?
Presentation 
Position   
Data  
Sampling   
Activity 
Self-selection  
Interaction  
Social  
Second order   

[Silberzahn et al., COS, Univ. of Virginia, 2015]
Professional Bias?
➔ 61 analysts, 29 teams: 20 yes and 9 no

If Systems Reflects Its Designers:
What we can/should do?
▪ Data
▪ Analyze for known and unknown biases, debias when possible/needed
▪ Recollect more data for difficult/sparse regions of the problem
▪ Delete attributes associated directly/indirectly with harmful bias
▪ Interaction
▪ Make sure that the user is aware of the biases all the time
▪ Give more control to the user
▪ Design and Implementation
• Let experts/colleagues/users contest every step of the process
▪ Evaluation
• Do not fool yourself!

Final Take-Home Message
▪ Systems are a mirror of us, the good, the bad and the ugly
▪ The Web amplifies everything, but always leaves traces
▪ We need to be aware of our own biases!
▪ We must be aware of the biases and contrarrest them to stop
the vicious bias cycle
▪ Plenty of open (research) problems!
Big Data of People is huge…..
….. but it is tiny compared to the future
Big Data of the Internet of Things (IoT)
No activity bias!

Thank You! Any Questions?
–
rbaeza@acm.org | http://www.baeza.cl/ | http://fairness-measures.org
Biased Questions?
ASIST 2012
Book of the
Year Award
(Biased Ad)

Thanks for Viewing the Slideshare!
–
Watch the Recording: https://youtube.com/session-example
Or
Contact us today to discover how Catalyst can deliver unparalleled SEO
results for your business. https://www.catalystdigital.com/

Keynote: Bias in Search and Recommender Systems

Recomendados

Recomendados

Mais conteúdo relacionado

Mais procurados

Mais procurados (20)

Semelhante a Keynote: Bias in Search and Recommender Systems

Semelhante a Keynote: Bias in Search and Recommender Systems (20)

Mais de Catalyst

Mais de Catalyst (20)

Último

Último (20)

Keynote: Bias in Search and Recommender Systems