Tutorial on Bias in Rec Sys @ UMAP2020

Hands on Data and Algorithmic
Bias in Recommender Systems
ACM UMAP2020
28th
Conference on User Modeling, Adaptation and Personalization
July 12, 2020 – ONLINE from GENOA

About us
2
Ludovico Boratto
Senior Research Scientist
Data Science and Big Data Analytics
EURECAT - Centre Tecnológic de Catalunya
Barcelona, Spain
ludovicoboratto.com
ludovico.boratto@acm.org
Mirko Marras
Postdoctoral Researcher
Department of Mathematics and Computer Science
University of Cagliari
Cagliari, Italy
mirkomarras.com
mirko.marras@unica.it
Hands on Data and Algorithmic Bias in Recommender SystemsBoratto and Marras

Learning objectives
● Raise awareness on the importance and the relevance of considering data and
algorithmic bias issues in recommendation
● Play with recommendation pipelines and conduct exploratory analysis aimed at
uncovering sources of bias along them
● Showcase approaches that mitigate bias along with the recommendation pipeline
and assess their inﬂuence on stakeholders
● Provide an overview on the current trends and challenges in bias-aware research
and identify new research directions
3Hands on Data and Algorithmic Bias in Recommender SystemsBoratto and Marras

Outline and scheduling
● 10:10 - 11:30 Session I: Foundations
○ 10:10 - 10:20 Recommendation Principles
○ 10:20 - 11:30 Data and Algorithmic Bias Fundamentals
● 11:30 - 12:00 Zoom Breakout Room + Q&A
● 12:00 - 13:10 Session II: Hands-on Case Studies
○ 12:00 - 12:15 Recommender Systems in Practice
○ 12:15 - 12:40 Investigation on Item Popularity Bias
○ 12:40 - 13:10 Investigation on Item Provider Fairness
● 13:10 - 13:20 Research Challenges and Emerging Opportunities
● 13:20 - 14:00 Open Discussion + Q&A
All times are displayed in conference local time (UTC+00:00)

What products could I buy?
Recommendation principles

What courses could I attend?

The problem

Recommender
System
A solution

Capitalizing on recommender systems
A recommender system suggests items that might be relevant for a user

The recommendation ranking task
12
● Given:
○ a set of consumers C = {c1
, c2
, ..., cM
}
○ a set of items I = {i1
, i2
, ..., iN
}
● Let R ⊆ R M ×N
be the consumer-item feedback matrix:
○ R(c,i) ≥ 0 if consumer c expressed interest in item i
○ R(c,i) = 0 otherwise
● The objective is to predict unobserved consumer-item feedback R(c,i) = f(c,i | θ) in R:
○ θ denotes model parameters
○ f denotes the function that maps model parameters to the predicted relevance
● Given a consumer c, items not rated by c are ranked by decreasing relevance:
i* = arg max f(c,j | θ)
j ∈ I Ic

Modes of optimization
13
● Pointwise optimization
point-wise approaches take a user-item pair and predict how relevant the item is for that user
● Pairwise optimization
pair-wise approaches digest a triplet of user, observed item, and unobserved item, and minimize
the cases when the unobserved item is more relevant than the observed item for that user
● Listwise optimization
list-wise approaches look at the entire list and build the optimal ordering for that user
Recommender systems in practice

Core recommendation techniques
Adapted from [Ricci et al. 2015]
14
Technique Background Input Process
Collaborative Ratings from C of items in I Ratings from c of items in I Identify users in C similar to c and
extrapolate from their preferences of i
Content-based Features of items in I Ratings from c of items in I Generate a classiﬁer that ﬁts c's rating
behavior and use it on i
Demographic Demographic information on C and
their ratings of items in I
Demographic information on c Identify users that are demographically
similar to c and extrapolate from their
preferences of i
Utility-based Features of items in I A utility function over items in I that
describes c's preferences
Apply the function to the items and
determine i's rank
Knowledge-based Features of items in I and knowledge of
how these items meet a user's need
A description of c's needs or interests Infer a match between i and c's needs or
interests

Core stakeholders in recommendation
[Abdollahpouri et al. 2020]
A recommendation stakeholder is any group or individual that can affect, or is affected
by, the delivery of recommendations to users
15
Consumers Providers System
C P S

Multi-sided recommendation aspects
16
Aspect Deﬁnition
Multi-stakeholder design A multistakeholder design process is one in which different recommendation stakeholder groups are
identiﬁed and consulted in the process of system design
Multi-stakeholder algorithm A multistakeholder recommendation algorithm takes into account the preferences of multiple parties when
generating recommendations, especially when these parties are on different sides of the recommendation
interaction
Multi-stakeholder evaluation A multistakeholder evaluation is one in which the quality of recommendations is assessed across multiple
groups of stakeholders, in addition to a point estimate over the full user population

A sample multi-sided scenario
17
Consumers
Students
Providers
Teachers
System
Online Course Platform

Data and algorithmic
bias fundamentals

Motivating example in music
[Mehrotra et al. 2018]
● People frequently listen to music online
● Ratings and frequencies often used to learn patterns
● 1/3 of users listen to at least 20% of unpopular artists
● Why are popular artists favoured?
● Why do users who tend to interact with niche artists
receive the worst recommendations?
Data and algorithmic bias fundamentals > Motivating examples

Motivating example in education
[Boratto et al. 2019]
● Online course platforms are receiving great attention
● Student's preferences learnt from ratings/enrolments
● The imbalance in popularity among courses reinforces
coverage and concentration biases of ranked courses
● Popularity bias could impede new courses to emerge
● The market could be dominated by a few teachers

Motivating example in social platforms
[Edizel et al. 2020]
● Reading users' stories is a common activity nowadays
● Users can vote stories up and down
● Gender attributes are not supported by Reddit
● Why recommender systems reinforce the imbalance
between genders while suggesting reddits?
● 95% (87%) of sub-reddits popular among females
(males) show imbalance reinforcement over genders

Motivating example in recruiting
[Singh et al. 2018]
22
● Recruiters rely more and more on automated systems
● Based on the job, "best" candidates are suggested
● Small differences in relevance can lead to large
differences in exposure among candidate groups
● Is this winner-take-all allocation of exposure fair, even
if the winner just has a tiny advantage in relevance?
● It might be fairer to distribute exposure proportional
to relevance, even if this leads to a drop in utility

Disclaimers
● We aim to focus on scientific literature that specifically consider recommender systems
● Pointers to representative scientific events on related concepts applied to ranking systems are given
● References discussed throughout the slides would not be exhaustive
● Refer to the extended bibliography attached to this tutorial for a more comprehensive list of papers
Data and algorithmic bias fundamentals

Related scientific venues and initiatives
Special issues in journals
including:
Scientific tutorials
including:
Dedicated workshops or tracks
including:
Papers in top-tier conferences
including:
● RecSys
● SIGIR
● The Web Conf
● TREC
● UMAP
● WSDM
● CIKM
● ECIR
● KDD
● FAccTRec @ RecSys 2018-2020
● RMSE & Impact RS @ RecSys 2019
● FACTS-IR @ SIGIR 2019
● FATES @ The Web Conf 2019
● FAIR-TREC @ TREC 2020
● FairUMAP @ UMAP 2018-2020
● DAB @ CIKM 2017-2019
● Bias @ ECIR 2020
● FAT-ML @ ICML 2014-2019
● Fairness and Discrimination in Retrieval and Recommendation @ SIGIR 2019 & RecSys 2019
● Learning to Rank in theory and practice: From Gradient Boosting to Neural Networks and Unbiased Learning @ SIGIR 2019
● Multi-stakeholder Recommendations: Case Studies, Methods and Challenges @ RecSys 2019
● Experimentation with fairness-aware recommendation using librec-auto @ FAT 2020
● Special Issue on Fair, Accountable, and Transparent Recommender Systems @ UMUAI
● Special Issue on Algorithmic Bias and Fairness in Search and Recommendation @ IPM
Data and algorithmic bias fundamentals > Scientific context

Perspectives impacted by bias
Economic
Legal
Social
Security
Technological
bias can introduce
disparate impacts
among providers,
influencing future success
and revenues
bias can affect core user's
rights that are regulated
by law, such as fairness
and discrimination
bias can reinforce
discrimination of certain
user's groups, including
ageism, sexism,
homophobia
bias can lead certain
groups of users or an
entire system to be more
vulnerable to attacks
(e.g., bribery)
bias can influence how
technologies progress and
can be amplified as the
algorithms evolve
Data and algorithmic bias fundamentals > Impacted perspectives

Law and rights impacted by bias
[Tolan et al. 2019]
● The right to non-discrimination, which can be undermined by inherent biases, is
embedded in the normative framework of the European Union, e.g.:
○ Explicit mentions of it can be found in Article 21 of the EU Charter of Fundamental Rights
○ Article 14 of the European Convention on Human Rights
○ Articles 18-25 of the Treaty on the Functioning of the European Union
● As an example, United Nations Sustainable Development Goal 4 aims also to "ensure
inclusive and equitable quality education and promote lifelong learning opportunities for all"
○ In Yao et al,. [2017], the authors observed that, in 2010, women accounted for only 18% of the
bachelor’s degrees awarded in computer science. The underrepresentation of women causes
historical rating data of computer-science courses to be dominated by men. Consequently, the
learned model may underestimate women’s preferences and be biased toward men.

Social aspects associated to bias
[Fabbri et al. 2020]
● Context: people recommendation in social networks, with users divided into groups based on gender
● Algorithm: Adamic-Adar, SALSA, ALS
● Findings: people recommenders produce disparate visibility on the two subgroups. Homophily plays a
key role in promoting or reducing visibility for different subgroups.
Lorenz Curves (inequality). Recommendations
introduce more inequality than the degree
distribution, and this inequality is stronger in
the minority class.

● Context: Sellers might attack the system by introducing a bias in the ratings
○ Attack goal: bribe the users to increase ratings and push item recommendations
● Algorithm: Novel hybrid KNN CF (each user is represented as a compressed string)
● Findings: proﬁtability associated to increasing the ratings is strongly reduced w.r.t. SV. Downgrading
the ratings of competitors is not proﬁtable with this approach, while it is with SVD. System is more
robust to attacks and more trustable by the users.
Security aspects undermined by bias
[Ramos et al. 2020]

Ethical aspects inﬂuenced by bias
[Bozdag 2013, Milano et al. 2020]
Content
recommendation of
inappropriate content
Opacity
black-box algorithms, uninformative
explanations, feedback effects
Privacy
unauthorised data collection, data
leaks, unauthorised inferences
Fairness
observation bias,
population imbalance
Autonomy and Identity
behavioural traps and encroachment
on sense of personal autonomy
Social
lack of exposure to contrasting
viewpoints, feedback effects
Data and algorithmic bias fundamentals > Inﬂuenced ethical aspects

Bias leading to inappropriate content
[Pantakar et al. 2019]
● Context: news recommender system who generates
awareness on biased news, to possibly avoid fake and
politically polarized news
● Algorithm: news clustering and bias score attached to
each news. Recommendation of similar, unbiased content
● Findings: live-user evaluation. The rankings generated by
the algorithm matches with the ones the users would
generate

Privacy of user representations
[Resheff et al. 2018]
● Context: user representations may be used to recover private user information such as gender and age
● Algorithms: privacy-adversarial framework to eliminate leakage of private information. An adversarial
component is appended to the model for each of the demographic variables to obfuscate, so that the
learned user representations are optimized to preclude predicting the variables
● Findings: privacy preserving recommendations, minimal overall adverse effect on recommender
performance, fairness of results (all knowledge of the attributes is scrubbed from the representations
used by the model)

Inﬂuence of bias on autonomy
[Arnold et al. 2018]
● Context: text recommender systems that support creative tasks
(domain: writing restaurant reviews). Do they exhibit
unintentional biases in the support that they offer? Do these
biases affect what people produce using these systems?
● Algorithms: contextual recommendations: (1) it selects the three
most likely next-word predictions, (2) it generates the most likely
phrase continuation for each word using beam search
● Findings: People who get recommended phrasal text entry
shortcuts that are skewed positive, write more positive reviews
than when presented with negative-skewed shortcuts

Objectives inﬂuenced by bias
[Kaminskas et al. 2017, Namatzadeh et al 2018, Singh et al. 2018]
Utility
Recommendation
Objectives
Novelty
Diversity
Coverage
Serendipity
the degree to which recommended
items are potentially useful and of
interest for the user
the degree of attention
received by (groups of)
items or providers
the degree to which the list has
valuable items not looked for and
generate surprise for the user
the degree to which the generated
recommendations cover the catalog
of available items
the degree to which the list of
retrieved items covers a broad area
of the information space
the degree to which items are unknown
by the user and/or are different from
what the user has seen before
Visibility &
Exposure
Data and algorithmic bias fundamentalsData and algorithmic bias fundamentals > Inﬂuenced objectives

Impact of bias on utility
[Fu et al. 2020]
● Context: study recommendation performance
according to the level of activity of users. Inactive
users are more susceptible to unsatisfactory
recommendations (insufﬁcient training data) +
recommendations are biased by the training records of
more active users.
● Algorithm: explainable CF + re-ranking to balance
predictions and group/individual fairness
● Findings: reduce disparity in utility while preserving
recommendation quality

Impact of bias on diversity
[Channamsetty and Ekstrand 2017, Lee and Hosanagar 2019]
● CF does not propagate users’ preferences for popularity and diversity into the
recommendations → lack of personalization
● Recommender systems lead to a decrease in sales diversity w.r.t. environments w/o
recommendations
● Recommenders can help individuals explore new products, but similar users end up
exploring the same kinds of products resulting in concentration bias at the
aggregate level

Trade-offs often come up
[Leonhardt et al. 2018, Boratto et al. 2020a]
● The introduction of diversity thanks to a
post-processing leads to an increasing disparity in
recommendation quality
● Debiasing NMF and BPR in terms of
popularity leads to a trade-off between
accuracy and beyond-accuracy metrics

Recourse and item availability
[Dean et al. 2020]
● Context: The amount of recourse available to a user is the percentage of unseen items that are
reachable (user-centric perspective). The availability of items in a recommender system is the
percentage of items that are reachable by some user (item-centric perspective).
● Algorithms: Linear preference models (SLIM and MF)
● Findings: unavailable items are systematically less popular than available items. Users with smaller
history lengths have more available recourse

Recommendation pipeline
39
Platform
Data
ModelRecommendations
Data Acquisition
and Storage
Data
Preparation
Model
Prediction
Recommendation
Delivering
Model
Evaluation
Pre-
Processed
Data
Model Setup
and Training
Data and algorithmic bias fundamentalsData and algorithmic bias fundamentals > Bias through the pipeline

Types of bias associated to users
[Olteanu et al. 2017]
● Population biases
differences in demographics or other user characteristics, between a population of users represented in a
dataset/platform and a target population
● Behavioral biases
differences in user behavior across platforms or contexts, or across users represented in different
datasets
● Content biases
behavioral biases that are expressed as lexical, syntactic, semantic, and structural differences in the
contents generated by users
● Linking biases
behavioral biases that are expressed as differences in the attributes of networks obtained from user
connections, interactions or activity
● Temporal biases
differences in populations or behaviors over time

Bias on items due to their popularity
[Jannach et al. 2014]
● Context: movies, books, hotels, and mobile games
● Algorithms: CB-Filtering, SlopeOne, User-KNN, Item-KNN, FM (MCMC), RfRec, Funk-SVD,
Koren-MF, FM (ALS), BPR
● Findings: techniques performing well on accuracy focus their recommendations on a tiny fraction of
the item spectrum or recommend mostly top sellers

To what degree popularity is good?
[Cañamares and Castells 2018]
● Context: movies
● Algorithms: User KNN, Item KNN, AvgRating, MF, Random, Pop
● Findings: effectiveness or ineffectiveness of popularity depends on the interplay of three main
variables: item relevance, item discovery by users, and the decision by users to interact with
discovered items. Authors identify the key probabilistic dependencies among these factors

Bias affecting item categories
[Guo and Dunson 2015, Lin et al. 2019]
● Items of different genres have different rating
values and different samples
● Bayesian multiplicative probit model to uncover
category-wise bias in ratings
● User preferences propagate differently in CF
recommendations, according to movie genre
and user gender
● Majority user group contributes with more
neighbors and inﬂuence predictions more
● SVD++ and BiasedMF dampen the preference
bias for movie genres for both men and women
● WRMF is well-calibrated for Sci-Fi/Crime for
both men and women but the behavior is
inconsistent for Action/Romance

Bias on ratings based on proximity
[Hu et al. 2014]
● Context: business recommendation (Yelp venues, e.g., restaurant, a shopping mall)
● Algorithms: MF + geographic inﬂuence
● Findings: there is a weak positive correlation between a business’s ratings and its neighbors’ ratings.
Geographical distance between a user and a business adversely affects the prediction accuracy.
Geographic inﬂuence helps improving prediction accuracy w.r.t. classic MF approaches

Biases conveyed by user's reviews
[Piramuthu et al. 2012, Xu et al. 2018, Dai et al. 2018, Vall et al. 2019]
● Sequential bias: the sequence in which reviews are written play an
appreciable role in how the reviews that follow later in the
sequence are written.
○ A theoretical model was devised, but real impact is left as future work
● Opinion bias: given a user–item pair, the opinion bias is defined as
the bias between rating and review. The rating matrix is filled with
a linear combination of the rating and the review sentiment
● Textual bias inspects how recommenders systems are influenced
by the fact that words may express different meanings in review
context
● Song order in sequence-aware recommendations: RNN-based
recommenders can learn patterns from song order, but they do not
help improving recommendation effectiveness

Time-related bias on local popularity
[Anelli et al. 2019]
● Context: popularity is a local concept and a function of time. Popular and recently rated items are
deemed as relevant for the user. Movies and toys recommendation.
● Algorithms: User KNN with the concept of precursor neighbors and a time decay
● Findings: Time-aware neighbors and local popularity lead to a comparable effectiveness (in terms of
NDCG w/ rime-independent rating order condition) + an improved efﬁciency

Types of bias in platforms
● Functional biases
biases that are a result of platform-speciﬁc mechanisms or affordances, that is, the possible actions within
each system or environment
● Normative biases
biases that are a result of written norms or expectations about unwritten norms describing acceptable
patterns of behavior on a given platform
● External biases
biases resulting from factors outside the platform, including considerations of socioeconomic status,
education, social pressure, privacy concerns, interests, language, personality, and culture
● Non-individual accounts
interactions on social platforms that are not produced by individuals, but by accounts representing
various types of organizations, or by automated agents

Bias in implicit/explicit feedback loop
[Hofmann et al. 2014]
● Context: how a recommender system’s evaluation
based on implicit feedback relates to rating-based
evaluation, and how evaluation outcomes may be
affected by bias in user behavior
● Findings:
○ implicit and explicit evaluation agree well when
assumptions agree well (e.g., precision@10 and CTR
with no-bias)
○ match between assumption on user behavior and
explicit evaluation matters – if assumptions are
violated, the wrong recommender can be preferred

● Context: characterize the impact of human-system feedback loop in the context of recommender
systems, demonstrating the unintended consequences of algorithmic confounding
● Findings:
○ the recommendation feedback loop
causes homogenization of user behavior
○ users experience losses in utility due
to homogenization effects
○ the feedback loop ampliﬁes the impact
of recommender systems on the
distribution of item consumption
Homogeneity in recommendation loop
[Chaney et al. 2018]

Anchoring preferences to suggestions
[Adomavicius et al 2013]
● Context: explore how consumer preferences are impacted by predictions of recommender systems
● Findings:
○ the rating presented by a recommender serves as an anchor for the consumer’s preference
○ viewers’ preference ratings can be signiﬁcantly inﬂuenced by the recommendation received
○ the effect is sensitive to the perceived reliability of a recommender system
○ the effect of anchoring is continuous and linear, operating over a range of system perturbations

Missing-not-at-random feedback
[Pradel et al. 2012]
● Context: study two major biases of the selection of
items, i.e., some items obtain more ratings than
others (popularity) and positive ratings are
observed more frequently than negative ratings
(positivity)
● Findings:
○ considering missing data as a form of negative
feedback during training may improve performances
○ ...but it can be misleading when testing, favoring
popularity more than user preferences

Misleading cues can bias user's views
[Elsweiler et al. 2017]
● Context: they explore the feasibility of substituting
meals that would typically be recommended to
users with similar, healthier dishes, investigating
how people perceive and select recipes
● Findings:
○ participants are unable to reliably identify which
recipe contains most fat due to their answers being
biased by lack of information
○ perception of fat content can be inﬂuenced by the
information available and, in some cases, misleading
cues (image or title) can bias a false impression

Sample of external bias
[Jahanbakhsh et al. 2020]
● Context: study how the ratings people receive on online labor
platforms are inﬂuenced by their performance, gender, their
rater’s gender, and displayed ratings from other raters
● Findings:
○ when the performance of paired workers was similar, low-performing
females were rated lower than their male counterparts
○ where there was a clear performance difference between paired workers,
low-performing females were preferred over a similarly-performing males
○ displaying an average rating from other raters made ratings more extreme,
resulting in high performing workers receiving signiﬁcantly higher ratings
and low performers lower ratings compared to absence of ratings

Bias on preference-consistent info
[Schwind et al. 2012]
● Context: when a diversity of viewpoints on controversial issues is available, learners
prefer information that is consistent with their prior preferences; so, they investigated
the role of two potential moderators (prior knowledge; cooperation vs. competition) on:
○ confirmation bias: the tendency to select more preference-consistent information
○ evaluation bias: the tendency to evaluate preference-consistent information as better
● Findings:
○ preference-inconsistent recommendations can be used to overcome this bias
○ preference-inconsistent recommendations reduced confirmation bias irrespective of prior
knowledge; evaluation bias was only reduced for participants with no prior knowledge
○ preference-inconsistent recommendations led to reduced confirmation bias under cooperation
and under competition; evaluation bias was only reduced under cooperation.

Decision biases in user's choices
[Teppan and Zanker 2015]
● Context: experimental analysis of the impact of different decision biases like decoy or position
effects, as well as risk aversion in positive decision frames on users' choice behavior
● Findings: strong dominance of risk aversion strategies and the need for awareness of these effects

Sources of bias in data collection
● Data acquisition
○ discouraging data collection by third parties
○ programmatic limitation of access to data (e.g., time, amount, size)
○ not all relevant data captured by the platform
○ opaque and unclear sampling strategies
● Data querying
○ limited expressiveness of APIs regarding information needs
○ different ways of operationalization of information by APIs
○ influence of keywords on datasets in keyword-based queries
● Data filtering
○ removal of outliers that are relevant for the analysis
○ bounding of analysis due to text filtering operations

Sources of bias in data preparation
● Data cleaning
○ data representation choices and default values
○ normalization procedures (e.g., based on geographical information)
● Data enrichment
○ subjective and noisy labels due to manual annotations
○ errors due to automatic annotation based on statistical or machine learning
● Data aggregation
○ lose of information due to high-level aggregation
○ spurious patterns of association when data is groups based on certain attributes

Sources of bias in model exploitation
● Qualitative analysis
○ data representation choices and default values
○ normalization procedures (e.g., based on geographical information)
● Descriptive analysis
○ research often relying on counting entities
○ inﬂuence of bias and confounders on correlation analysis
● Inference analysis
○ variations of performance across and within datasets
○ deﬁnition of target variables, class labels, or data representations
○ effect of the objective function to the inference task
● Observational analysis
○ peer effects due to platform affordances and conventions
○ selection bias and how treatment effects on results generalizability

Sources of bias in model evaluation
[Olteanu et al. 2017, Bellogin et al. 2017]
● Evaluation data selection
○ imbalances of data samples due to their popularity
○ sensitivity to the ratio of the test ratings versus the added non-relevant items
● Metrics selection
○ inﬂuence of choice of metrics on research study takeaways
○ accounting domain impact throughout performance assessment
● Result assessment and interpretation
○ traces and patterns changing with context
○ going beyond studies evaluated on a single dataset or method

Facing popularity bias in evaluation
[Bellogin et al. 2017]
● a percentile-based approach consists in:
○ dividing items in m popularity percentiles
○ breaking down the computation of
accuracy by such percentiles
○ averaging the m obtained values
● a uniform test approach consists in:
○ formation of data splits where all items
have the same amount of test ratings
○ picking a set T of candidate items and a
number g of test ratings per item

Random decoys in evaluation
[Ekstrand et al. 2017] [Other readings: Lim et al. 2015, Yang et al. 2018, Carraro et al. 2020]
● Context: examine the random decoys protocol, where the
candidate set consists of the test set items plus a
randomly-selected set of N decoy items
● Findings:
○ the distribution of items goodness required to avoid
misclassiﬁed decoys with reasonable probability is
unreasonable
○ there is a serious discrepancy between theoretical
and observed behavior of the random decoy
strategy with respect to popularity bias

Error bias in evaluation
[Tian et al. 2020]
● Context: offline evaluation cannot accurately
assess novel, relevant recommendations,
because the most novel items are missing from
the data and cannot be judged as relevant
● Findings:
○ missing data in the observation process causes
the evaluation to mis-estimate metric values
○ substantial breakthroughs in recommendation
quality will be difficult to be assessed offline

Biases that lead to discrimination
[Mehrabi et al. 2019]
Direct
Discrimination
direct discrimination happens when
protected attributes of groups or
individuals explicitly result in
non-favorable outcomes toward them
Indirect
Discrimination
individuals appear to be treated based on neutral
and non-protected attributes; however, protected
groups or individuals get to be treated unjustly as a
result of implicit effects from protected attributes
Data and algorithmic bias fundamentalsData and algorithmic bias fundamentalsData and algorithmic bias fundamentals > Discrimination

Granularity of discrimination
Individual
Discrimination
when a system gives unfairly different
predictions to individuals who are
considered similar for that task
Group
Discrimination
when a system systematically treats
individuals who belong to different
groups unfairly
Sub-group
Discrimination
when a system systematically
discriminate individuals over a
large collection of subgroups

Contextualizing to recommendation
● Individual and sub-group discrimination are very challenging to assess: user similarity
at individual level should be on intrinsic properties that characterize the users and not
based on behavioral aspects
● No study so far in the RecSys domain
● Advances in ranking systems, where recommendation approaches can draw from:
[Zehlike et al. 2017, Celis et al. 2017, Biega et al. 2018, Lahoti et al. 2019, Yada et al. 2019, Singh and Joachims
2019, Kuhlman et al. 2019, Zehlike and Castillo 2020, Ramos and Boratto 2020a, Diaz et al. 2020]
● For much more, see the "Fairness & Discrimination in Retrieval & Recommendation"
tutorial at https://fair-ia.ekstrandom.net/sigir2019-slides.pdf

Types of disparity in groups
Disparate
Treatment
when members of different
groups are treated differently
Disparate
Impact
when members of different groups
obtain different outcomes
Disparate
Mistreatment
when members of different groups
have different error rates

Contextualizing disparities
● Recommender systems usually do not receive as input any sensitive attribute of the
user, so disparate treatment is usually not considered
● Disparate impact usually does not affect recommender systems: two users of
different genders with the same ratings for the same items, usually receive the
same recommendations
● Disparate mistreatment means that groups of users with different sensitive
attributes receive different recommendation quality. Known as consumer fairness
in the RecSys domain (presented later)

Deﬁnitions of fairness
Equalized Odds
an algorithm is fair if the groups have
equal rates for true positives and false
positives
Fairness through
Awareness
an algorithm is fair if it gives similar
predictions to similar individuals
Equal Opportunity
an algorithm is fair if the groups have
equal true positive rates
Fairness through
Unawareness
an algorithm is fair as long as any
protected attributes is not explicitly
used in the decision-making process
Demographic Parity
an algorithm is fair if the likelihood of a
positive outcome should be the same
regardless of the group
Equality of Treatment
an algorithm is fair if the ratio of false
negatives and false positives is the
same for all the groups

Fairness metrics for ranked outputs
[Yang and Stoyanovich 2017, Farnadi et al. 2018, Sonboli et al. 2019]
● Normalized discounted difference (rND) computes the difference in the proportion of
members of the protected group at top-i and in the over-all population
● Normalized discounted KL-divergence (rKL) measures the expectation of the logarithmic
difference between two discrete probability distributions
● Local fairness supports the fact that fairness may be local and the identiﬁcation of protected
groups is only possible through consideration of local conditions
● Non-parity unfairness measures the absolute difference between the overall average ratings
of users belonging to the unprotected and the protected group
● Value unfairness measures the inconsistency in signed estimation error across the protected
and unprotected groups (becomes larger when predictions for one group are overestimated)
● Absolute unfairness measures the inconsistency in absolute estimation error across user
groups (becomes large if one group consistently receives more accurate recommendations)

Multi-sided fairness
[Burke et al. 2017]
Consumer-Provider Fairness
It might be needed that the platform
guarantees fairness for both
consumers and providers, e.g.,:
● people matching
● property/business
recommendation
● user skills and job matching
● and so on...
Consumer Fairness
We talk about unfairness for consumers
when their experience in the platform
differs:
● in terms of service effectiveness
(results’ relevance, user
satisfaction)
● resulting outcomes (exposure to
lower-paying job offers)
● participation costs (differential
privacy risks)
Provider Fairness
Providers experience unfairness when a
platform/service creates:
● different opportunities for their
items to be consumed
● different visibility or exposure in
the ranking
● different participation costs

Impact of bias on consumer fairness
[Ekstrand et al. 2018a]
● Context: they investigate whether demographic
groups obtain different utility from recommender
systems in LastFM and Movielens 1M datasets
● Algorithms: Popular, Mean Item-Item, User-User,
FunkSVD
● Findings: ML1M & LFM1K have statistically-signiﬁcant
differences between gender groups, and LFM360K has
signiﬁcant differences between age brackets

Impact of bias on consumer fairness
[Kowald et al. 2020]
● Context: investigate three user groups from Last.fm based on how much their listening preferences
deviate from the most popular music, among all Last.fm users: (i) low-mainstream users, (ii)
medium-mainstream users, and (iii) high-mainstream users
● Findings: low-mainstreaminess group signiﬁcantly receives the worst recommendations

Impact of bias on provider fairness
[Ekstrand et al. 2018b]
● Context: examine the response of collaborative ﬁltering algorithms to the distribution of their input
data, with respect to content creator’s gender
● Findings: matrix factorization produced reliably male-biased recommendations, while nearest-neighbor
and hierarchical Poisson factorization techniques were closer to the user proﬁle tendency while being
less diffuse than their inputs

Bias-aware process pipeline
76
IDENTIFY PRODUCT GOALS
● What are you trying to achieve?
● For what population of people?
● What metrics are you tacking?
MITIGATE ISSUES
● Does data include enough minority samples?
● Do our proxies measure what we think they do?
● Does the bias notion capture customer needs?
IDENTIFY STAKEHOLDERS
● Who has a stake in this product?
● Who might be harmed?
● How?
DEVELOP AND ANALYZE THE SYSTEM
● How well the system matches product goals?
● To what degree bias is still present?
● How decisions impact on each stakeholder?
DEFINE A BIAS NOTION
● What type of bias? At what point?
● What distributions?
Bias-aware
process pipeline
Data and algorithmic bias fundamentalsData and algorithmic bias fundamentalsData and algorithmic bias fundamentals > Mitigation design

Techniques for bias treatment
77
Pre-processing
before model training
In-processing
during model training
Post-processing
after model training
Pre-processing techniques try to
transform the data so that the
bias is mitigated. If the algorithm
is allowed to modify the training
data, pre-processing can be used
In-processing techniques try to
modify learning algorithms to
mitigate bias during training process.
If it is allowed to change the learning
procedure, in-processing can be used
Post-processing is performed by
re-ranking items of the lists obtained
after model training. If the algorithm
can treat the learned model as a black
box, post-processing can be used

Sample of pre-processing treatment
[Rastegarpanah et al. 2019]
● Idea: rather than transforming the system’s input
data, they investigate whether simply augmenting
the input with additional data can improve the
social desirability of the recommendations
● Algorithms: MF family of algorithms
● Findings: the small amounts of antidote data
(typically on the order of 1% new users) can
generate a dramatic improvement (on the order of
50%) in the polarization or the fairness of the
system’s recommendations

Sample of in-processing treatment
[Beutel et al. 2019]
● Idea: they propose a metric for measuring fairness
based on pairwise comparisons and devise a
correlation-based regularization approach to improve
model performance for the given fairness metric
● Algorithms: learning-to-rank (e.g., point- and pair-wise)
● Findings: while the regularization decreases the
pairwise accuracy of the non-subgroup items, it closes
the gap in the inter-group pairwise fairness, resulting in
only a 2.6% advantage for non-subgroup items in
inter-group pairwise fairness, down from 35.6%

Sample of in-processing treatment
● Idea: identify a regularization component of the
objective to be minimized when the distribution
of recommendations achieves a 50/50 balance
between medium-tail and short-head items. We
intend to generalize this objective in future work.
● Algorithms: RankALS (i.e., pair-wise learning)
● Findings: it is possible to model the trade-off
between long-tail catalog coverage and ranking
accuracy as a multi-objective optimization
problem based on a dissimilarity matrix

Sample of post-processing treatment
[Liu et al. 2019b]
● Idea: combining of a personalization-induced term
and a fairness-induced term, promoting the loans that
belong to currently uncovered borrower groups
● Algorithms: RankSGD, UserKNN, WRMF, Maxent
● Findings: they ﬁnd a balance between the two terms,
where nDCG still remains at a high level after the
re-ranking, while fairness of the recommendation is
signiﬁcantly improved, as loans belonging to
less-popular groups are promoted.

Other treatments against popularity
among others...
● Treatments that manipulate interactions before training a model:
○ sample tuples (u,i, j) where i is less popular than j for pair-wise learning [Jannach et al. 2014]
○ remove popular items, simulating situations in which these items are missing [Cremonesi et al. 2014]
○ detect and fix noisy ratings by characterizing items and users by their profiles [Toledo et al. 2015]
● Treatments that regularize the loss function score during training:
○ a regularization that balance recommendation accuracy and intra-list diversity [Abdollahpouri et al. 2017]
○ a regularization that minimizes the correlation between accuracy and item popularity [Boratto et al. 2020]
● Treatments that re-rank items after model training:
○ two-way aggregation of direct and reversed rank results (to improve coverage and accuracy) [Dong et al. 2020]
○ a re-ranking that suggests first items from unseen providers (to improve coverage) [Burke et al. 2016]
○ a re-ranking score that balances predicted rating with the inverse of popularity [Abdollahpouri et al. 2018]
○ a re-ranking that includes long-tail items the user might like [Abdollahpouri et al. 2019]

Other treatments against C-fairness
among others...
● Strategies that manipulate data of (groups of) individuals before training:
○ add new users who rate existing items to minimize polarization or improve fairness [Rastegarpanah et al. 2019]
● Strategies that introduce regularization or constraints during training:
○ create a balanced neighborhood from protected and unprotected classes [Burke et al. 2018]
○ objective functions pushing independence between predicted ratings and sensitive attributes [Kamishima et al. 2018]
○ a fairness-aware model that isolates and extracts sensitive information from latent factor matrices [Zhu et al. 2018]
○ a probabilistic programming approach for building fair hybrid recommender systems [Farnadi et al. 2018]
● Strategies that require a re-ranking of the items after training:
○ a heuristic re-ranking to mitigate unfairness in explainable recommendation over knowledge graphs [Fu et al. 2020]

Treatments for C-fairness of groups
among others...
● Stratigi et al. 2017:
○ improving the opportunities that patients have to inform themselves online about diseases and possible treatments
○ identify the correct set of similar users for a user in question, considering health information and ratings
○ a fairness-aware recommendation model of items that are relevant and satisfy the majority of the group members
● Lin et al. 2017:
○ maximize the satisfaction of each group member while minimizing the unfairness between them
○ introduce two concepts of social welfare and fairness, modeling overall utilities and balance between group members
○ an optimization framework for fairness-aware group recommendation from the perspective of Pareto Efficiency
● Serbos et al. 2017:
○ recommending packages of items to groups of users, e.g., recommending vacation packages to groups of tourists
○ fair in the sense that every group member is satisfied by a sufficient number of items in the package
○ propose greedy algorithms that find approximate solutions to meet a better trade-off within reasonable time

Other treatments against P-fairness
among others...
● Strategies involving some degree of pre-processing:
○ a mitigation approach that relies on tailored upsampling in pre-processing of interactions involving minority
groups of providers [Boratto et al. 2020b]
● Strategies that change or regularize the algorithm learning process:
○ the concept of balanced neighborhood from protected and unprotected class can be applied also to improve
fairness across providers [Burke et al. 2018]
○ a correlation-based regularization that minimizes the correlation between the residual between the clicked
and unclicked item and the group membership of the clicked item [Beutel et al. 2019]
● Strategies that re-rank items with a post-processing procedure:
○ a re-ranking to improve exposure distribution across creators, controlling divergence between the desired
distribution and the actual obtained distribution of exposure [Modani et al. 2017]
○ iteratively generate the ranking list by trading off between accuracy and the coverage of the providers based
on the adaptation of the xQuad algorithm [Liu et al. 2019b]

Treatments with a multi-sided focus
among others...
● In a speed-dating domain:
○ in a speed-dating context, a multi-dimensional utility framework which analyzes the relationship between
utilities and recommendation performance, achieving a trade-off [Zheng et al. 2018]
○ an approach to rerank the recommendation list by optimizing (1) the disparity of service; (2) the similarity of
mutual preference; (3) the equilibrium of demand and supply [Xia et al. 2019]
● Or in a generic multi-sided marketplace:
○ an integer linear programming-based optimization to deploy changes incrementally in steps, ensuring smooth
transition of item exposure and a minimum loss in utility [Patro et al. 2019]
○ an algorithm guarantees at least maximin share of exposure for most of the producers and envy-free up to one
good fairness for every customer [Patro et al. 2020]

Treatments against other biases (1)
● Biases related to how items are sampled, positioned, and/or selected, e.g.:
○ connect recommendation to causal inference from experimental and observational data [Schnabel et al. 2016]
○ integrate imputed errors and propensities, for alleviating the effect of propensity variance [Wang et al. 2019]
○ manage the spiral of silence effect, i.e., users are likely to rate if they perceive a support by the dominant opinion [Liu D.
et al. 2019a]
○ estimate item frequency from a data stream, subject to vocabulary and distribution shifts [Yi et al. 2019]
○ model position-bias ofﬂine and conduct online inference without position information [Guo et al. 2019]
○ off-policy correction to learn from feedback given by an ensemble of prior model policies [Chen et al. 2019a]
○ a clipped estimator to improve the bias-variance trade-off than w.r.t. unbiased estimator [Saito et al. 2020]
○ a counterfactual approach which accounts for selection and position bias jointly [Ovaisi et al. 2020]
○ a two-stage off-policy that takes the ranker model into account while training the candidate model [Ma et al. 2020]

Treatments against other biases (2)
● Biases associated to how items reach their audience:
○ a novel probabilistic method for weighted sampling of k neighbors that considers the similarity levels
between the target user (or item) and the candidate neighbors [Adamopoulos et al. 2014]
○ a target customer re-ranking algorithm to adjust the population distribution and composition in the
top-k target customers of an item while maintaining recommendation quality [Zhao et al. 2020]
● Biases associated to reviews and textual opinions, e.g.:
○ a sentiment classification scoring method, which employs dual attention vectors to predict the users’
sentiment scores of their reviews , to catch opinion bias and enhance user-item matrix [Xu et al. 2018]
○ a hybrid model that integrates modified-sied information related to textual bias and rating bias in
matrix factorization, getting a specific word representation for each item review [Dai et al. 2018]
● Biases associated to how items are marketed, e.g.:
○ a fairness-aware framework to address market imbalance bias by calibrating the parity of prediction
errors across different market segments [Wan et al. 2020]

Mitigations against other biases (3)
● Biases associated to social trust and influence:
○ a mitigation using polynomial regression and a Bayesian information criterion to predict ratings less influenced
by the tendency to conform to the perceived “norm” in a community [Krishnan et al. 2014]
○ clustering user-item space to discover rating bubbles derived from the theory of social bias, i.e., existing ratings
indirectly influences the users' opinion to follow the herd instinct [Divyaa et al. 2019]
○ a matrix completion algorithm that performs hybrid memory-based collaborative filtering, improving how the
bribery effect is managed and how the system is robust against bribery [Ramos et al. 2020b]
● Biases related to the interactions of users over time, e.g.:
○ an historical influence-aware latent factor model to capture and mitigate historical distortions in each single
rating under the assimilate-contrast theory: users conform to historical ratings if historical ratings are not far
from the product quality (assimilation), while users deviate from historical ratings if historical ratings are
significantly different from the product quality (contrast) [Zhang et al. 2018]
○ an unbiased loss using inverse propensity weighting, that includes the recency propensity of item x at time t, to
be used in point-wise learning to rank [Chen et al. 2019b]

BREAK
We resume at 12:00 UTC+00:00
Hands on Data and Algorithmic
Bias in Recommender Systems
ACM UMAP2020: 28th
Conference on User Modeling, Adaptation and Personalization
July 12, 2020 – ONLINE from GENOA

Recommender Systems
in Practice

Steps of this hands on
92
1
Data Load
We load data from publicly
available datasets, specifically
focusing on Movielens 1M
(movies)
Data Pre-Processing
We process data to be fed
into the model and we prepare
training samples, focusing on
pairwise data
2
Model Definition and Train
We define the architecture of
the model, setup the training
parameters and run the model
training process
3
Relevance Computation
Given a pre-trained model, we
compute the user-item
relevance scores across all
the user-item pairs
4
Model Evaluation
We compute accuracy and
beyond-accuracy metrics,
such as coverage, novelty, and
diversity
5
http://bit.ly/BiasInRecSysTutorial-NB1-UMAP2020

Disclaimers
● In this tutorial, we do not aim to show how to ﬁne-tune algorithms
● Due to the time constraints, we decided to reduce the optimization part
● The pre-trained models do not represent ﬁne-tuned baselines
● The goal is to get familiar with an environment where it is easier to control the whole recsys process

Case Study I
Item Popularity Bias

95
1
Model Exploration
We consider on data and
models introduced in the first
hands on to inspect of
popularity impacts on visibility
and exposure of items
Mitigation Setup
We arrange a representative
set of mitigation strategies
against popularity bias in pre-,
in- and post-processing
2
Mitigation Running
We run the mitigation
procedure, inspecting how the
optimization processes
influences popularity values
3
Model Re-Evaluation
We re-run the evaluation of
the first hands on to highlight
how disparities among
popular and unpopular items
are reduced
4
Impact Assessment
We interpret the results
obtained during evaluation in
order to envision how
stakeholders are impacted
5
N
Investigation on item popularity bias

Case Study II
Item Provider Fairness

97
1
Model Exploration
We consider on data and
models introduced in the first
hands on to inspect of
popularity impacts on visibility
and exposure of providers
Mitigation Setup
We arrange a representative
set of mitigation strategies
against provider unfairness in
pre-, in- and post-processing
2
Mitigation Running
We run the mitigation
procedure, inspecting how the
optimization processes
influences provider fairness
3
Model Re-Evaluation
We re-run the evaluation of
the first hands on to highlight
how disparities among
providers are reduced
4
Impact Assessment
We interpret the results
obtained during evaluation in
order to envision how
stakeholders are impacted
5
Investigation on item provider fairness

Research challenges and
emerging opportunities

Contextual challenges
Research challenges and emerging opportunities
● Different stakeholders have different (and possibly conﬂicting) needs. How can recommender systems
account for them?
● Multi-disciplinary approaches to go beyond recommendation algorithms (e.g., to link justice and
fairness)
● Synthesizing a deﬁnition of bias or fairness is challenging
● Creating a common vocabulary to recognize different types of bias and unfairness and advance as a
community
● Data to characterize bias phenomena with enough depth is lacking (especially for sensitive attributes)
● There are forms of bias on the Web that have not been studied in the recommendation literature

Operational challenges
● Measuring and operationalizing a deﬁnition of bias or fairness. How can we optimize a recommender
system for it?
● Can we mitigate multiple forms of bias at the same time?
● Slight changes throughout the pipeline can make a huge difference on impact
● Research and development should be more focused on the real world application
● When mitigating bias we usually trade for other qualities. How can we mitigate bias without
compromising recommendation quality?

Bridging offline and online evaluation
● What if we do not have the sensitive attributes in the collected data?
● How should we select an approach with respect to another (e.g., equity vs equality)?
● How to identify harms in the considered context?
● Will the chosen offline metrics and experiments lead to the desired results online?
● How to inspect whether data generation and collection methods are appropriate?
● How could we take into account both bias goals and efficiency in the real world?

To what extent are search and
recommendation algorithms
getting closer to each other?

Resources from this tutorial
● Tutorial website: http://bit.ly/BiasInRecSysTutorial-UMAP2020
● Github repository: http://bit.ly/BiasInRecSysTutorial-Github-UMAP2020
● Jupyter notebooks:
○ Neural recommender system train and test: http://bit.ly/BiasInRecSysTutorial-NB1-UMAP2020
○ Investigation on item popularity bias: http://bit.ly/BiasInRecSysTutorial-NB2-UMAP2020
○ Investigation on item provider fairness: http://bit.ly/BiasInRecSysTutorial-NB3-UMAP2020

References #1
1. Himan Abdollahpouri, Gediminas Adomavicius, Robin Burke, Ido Guy, Dietmar Jannach, Toshihiro Kamishima, Jan Krasnodebski, Luiz Augusto
Pizzato: Multistakeholder recommendation: Survey and research directions. User Model. User Adapt. Interact. 30(1): 127-158 (2020).
2. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher: Managing Popularity Bias in Recommender Systems with Personalized Re-Ranking.
FLAIRS Conference 2019: 413-418 (2019).
3. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher. Popularity-Aware Item Weighting for Long-Tail Recommendation. arXiv preprint
arXiv:1802.05382 (2018).
4. Himan Abdollahpouri, Robin Burke, Bamshad Mobasher: Controlling Popularity Bias in Learning-to-Rank Recommendation. RecSys 2017:
42-46 (2017).
5. Panagiotis Adamopoulos, Alexander Tuzhilin: On over-specialization and concentration bias of recommendations: probabilistic neighborhood
selection in collaborative ﬁltering systems. RecSys 2014: 153-160 (2014).
6. Gediminas Adomavicius, Jesse C. Bockstedt, Shawn P. Curley, Jingjing Zhang: Do Recommender Systems Manipulate Consumer Preferences? A
Study of Anchoring Effects. Inf. Syst. Res. 24(4): 956-975 (2013).
7. Vito Walter Anelli, Tommaso Di Noia, Eugenio Di Sciascio, Azzurra Ragone, Joseph Trotta: Local Popularity and Time in top-N Recommendation.
ECIR (1) 2019: 861-868 (2019).
8. Kenneth C. Arnold, Krysta Chauncey, Krzysztof Z. Gajos: Sentiment Bias in Predictive Text Recommendations Results in Biased Writing.
Graphics Interface 2018: 42-49 (2018).

References #2
9. Alejandro Bellogín, Pablo Castells, Iván Cantador: Statistical biases in Information Retrieval metrics for recommender systems. Inf. Retr. J.
20(6): 606-634 (2017).
10. Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H. Chi, Cristos Goodrow: Fairness in
Recommendation Ranking through Pairwise Comparisons. KDD 2019: 2212-2220
11. Asia J. Biega, Krishna P. Gummadi, Gerhard Weikum: Equity of Attention: Amortizing Individual Fairness in Rankings. SIGIR 2018: 405-414
12. Ludovico Boratto, Gianni Fenu, Mirko Marras: The Effect of Algorithmic Bias on Recommender Systems for Massive Open Online Courses.
ECIR (1) 2019: 457-472 (2019).
13. Ludovico Boratto, Gianni Fenu, Mirko Marras: Connecting User and Item Perspectives in Popularity Debiasing for Collaborative
Recommendation. CoRR abs/2006.04275 (2020a).
14. Ludovico Boratto, Gianni Fenu, Mirko Marras: Interplay between Upsampling and Regularization for Provider Fairness in Recommender
Systems. CoRR abs/2006.04279 (2020b).
15. Engin Bozdag: Bias in algorithmic ﬁltering and personalization. Ethics Inf Technol 15, 209–227 (2013).
16. Robin Burke. Multisided fairness for recommendation. arXiv preprint arXiv:1707.00093 (2017).
17. Robin Burke, Nasim Sonboli, Aldo Ordonez-Gauger: Balanced Neighborhoods for Multi-sided Fairness in Recommendation. FAT 2018: 202-214
(2018)
18. Robin D. Burke, Himan Abdollahpouri, Bamshad Mobasher, Trinadh Gupta: Towards Multi-Stakeholder Utility Evaluation of Recommender
Systems. UMAP Extended Proceedings (2016).

References #3
19. Rocío Cañamares, Pablo Castells: Should I Follow the Crowd?: A Probabilistic Analysis of the Effectiveness of Popularity in Recommender
Systems. SIGIR 2018: 415-424
20. Diego Carraro, Derek Bridge: Debiased ofﬂine evaluation of recommender systems: a weighted-sampling approach. SAC 2020: 1435-1442
(2020).
21. L. Elisa Celis, Damian Straszak, Nisheeth K. Vishnoi: Ranking with Fairness Constraints. ICALP 2018: 28:1-28:15 (2017).
22. Roberto Centeno, Ramón Hermoso, Maria Fasli: On the inaccuracy of numerical ratings: dealing with biased opinions in social networks. Inf.
Syst. Frontiers 17(4): 809-825 (2015).
23. Allison J. B. Chaney, Brandon M. Stewart, Barbara E. Engelhardt: How algorithmic confounding in recommendation systems increases
homogeneity and decreases utility. RecSys 2018: 224-232 (2018).
24. Sushma Channamsetty, Michael D. Ekstrand: Recommender Response to Diversity and Popularity Bias in User Proﬁles. FLAIRS Conference
2017: 657-660 (2017).
25. Minmin Chen, Alex Beutel, Paul Covington, Sagar Jain, Francois Belletti, Ed H. Chi: Top-K Off-Policy Correction for a REINFORCE
Recommender System. WSDM 2019: 456-464 (2019a).
26. Ruey-Cheng Chen, Qingyao Ai, Gaya Jayasinghe, W. Bruce Croft: Correcting for Recency Bias in Job Recommendation. CIKM 2019: 2185-2188
(2019b).
27. Paolo Cremonesi, Franca Garzotto, Roberto Pagano, Massimo Quadrana: Recommending without short head. WWW (Companion Volume)
2014: 245-246 (2014).

References #4
28. Jiao Dai, Mingming Li, Songlin Hu, Jizhong Han: A Hybrid Model Based on the Rating Bias and Textual Bias for Recommender Systems. ICONIP
(2) 2018: 203-214 (2018).
29. Sarah Dean, Sarah Rich, Benjamin Recht: Recommendations and user agency: the reachability of collaboratively-ﬁltered information. FAT*
2020: 436-445 (2020).
30. Fernando Diaz, Bhaskar Mitra, Michael D. Ekstrand, Asia J. Biega, Ben Carterette: Evaluating Stochastic Rankings with Expected Exposure.
CoRR abs/2004.13157 (2020).
31. Divyaa L. R., Nargis Pervin: Towards generating scalable personalized recommendations: Integrating social trust, social bias, and geo-spatial
clustering. Decis. Support Syst. 122 (2019).
32. Qiang Dong, Quan Yuan, Yang-Bo Shi: Alleviating the recommendation bias via rank aggregation. Physica A: Statistical Mechanics and its
Applications, 534, 122073. (2019).
33. Bora Edizel, Francesco Bonchi, Sara Hajian, André Panisson, Tamir Tassa: FaiRecSys: mitigating algorithmic bias in recommender systems. Int. J.
Data Sci. Anal. 9(2): 197-213 (2020)
34. David Elsweiler, Christoph Trattner, Morgan Harvey: Exploiting Food Choice Biases for Healthier Recipe Recommendation. SIGIR 2017:
575-584 (2017).
35. Michael D. Ekstrand, Mucun Tian, Ion Madrazo Azpiazu, Jennifer D. Ekstrand, Oghenemaro Anuyah, David McNeill, Maria Soledad Pera: All The
Cool Kids, How Do They Fit In?: Popularity and Demographic Biases in Recommender Evaluation and Effectiveness. FAT 2018: 172-186
(2018a).

References #5
36. Michael D. Ekstrand, Mucun Tian, Mohammed R. Imran Kazi, Hoda Mehrpouyan, Daniel Kluver: Exploring author gender in book rating and
recommendation. RecSys 2018: 242-250 (2018b)
37. Michael D. Ekstrand, Vaibhav Mahant: Sturgeon and the Cool Kids: Problems with Random Decoys for Top-N Recommender Evaluation. FLAIRS
Conference 2017: 639-644 (2017).
38. Francesco Fabbri, Francesco Bonchi, Ludovico Boratto, Carlos Castillo: The Effect of Homophily on Disparate Visibility of Minorities in People
Recommender Systems. ICWSM 2020: 165-175 (2020).
39. Golnoosh Farnadi, Pigi Kouki, Spencer K. Thompson, Sriram Srinivasan, Lise Getoor: A Fairness-aware Hybrid Recommender System. CoRR
abs/1809.09030 (2018).
40. Zuohui Fu, Yikun Xian, Ruoyuan Gao, Jieyu Zhao, Qiaoying Huang, Yingqiang Ge, Shuyuan Xu, Shijie Geng, Chirag Shah, Yongfeng Zhang, Gerard
de Melo: Fairness-Aware Explainable Recommendation over Knowledge Graphs. CoRR abs/2006.02046 (2020).
41. Huifeng Guo, Jinkai Yu, Qing Liu, Ruiming Tang, Yuzhou Zhang: PAL: a position-bias aware learning framework for CTR prediction in live
recommender systems. RecSys 2019: 452-456 (2019).
42. Fangjian Guo, David B. Dunson: Uncovering Systematic Bias in Ratings across Categories: a Bayesian Approach. RecSys 2015: 317-320 (2015).
43. Katja Hofmann, Anne Schuth, Alejandro Bellogín, Maarten de Rijke: Effects of Position Bias on Click-Based Recommender Evaluation. ECIR
2014: 624-630 (2014).
44. Longke Hu, Aixin Sun, Yong Liu: Your neighbors affect your ratings: on geographical neighborhood inﬂuence to rating prediction. SIGIR 2014:
345-354 (2014).

References #6
45. Farnaz Jahanbakhsh, Justin Cranshaw, Scott Counts, Walter S. Lasecki, Kori Inkpen: An Experimental Study of Bias in Platform Worker Ratings:
The Role of Performance Quality and Gender. CHI 2020: 1-13 (2020).
46. Dietmar Jannach, Lukas Lerche, Iman Kamehkhosh, Michael Jugovac: What recommenders recommend: an analysis of recommendation biases
and possible countermeasures. User Model. User Adapt. Interact. 25(5): 427-491 (2015).
47. Marius Kaminskas, Derek Bridge: Diversity, Serendipity, Novelty, and Coverage: A Survey and Empirical Analysis of Beyond-Accuracy
Objectives in Recommender Systems. ACM Trans. Interact. Intell. Syst. 7(1): 2:1-2:42 (2017).
48. Toshihiro Kamishima, Shotaro Akaho, Hideki Asoh, Jun Sakuma: Recommendation Independence. FAT 2018: 187-201 (2018).
49. Dominik Kowald, Markus Schedl, Elisabeth Lex: The Unfairness of Popularity Bias in Music Recommendation: A Reproducibility Study. ECIR (2)
2020: 35-42 (2020).
50. Sanjay Krishnan, Jay Patel, Michael J. Franklin, Ken Goldberg: A methodology for learning, analyzing, and mitigating social inﬂuence bias in
recommender systems. RecSys 2014: 137-144 (2014).
51. Caitlin Kuhlman, MaryAnn Van Valkenburg, Elke A. Rundensteiner: FARE: Diagnostics for Fair Ranking using Pairwise Error Metrics. WWW
2019: 2936-2942 (2019).
52. Preethi Lahoti, Krishna P. Gummadi, Gerhard Weikum: iFair: Learning Individually Fair Data Representations for Algorithmic
Decision Making. ICDE 2019: 1334-1345

References #7
53. Dokyun Lee, Kartik Hosanagar: How Do Recommender Systems Affect Sales Diversity? A Cross-Category Investigation via Randomized Field
Experiment. Inf. Syst. Res. 30(1): 239-259 (2019).
54. Jurek Leonhardt, Avishek Anand, Megha Khosla: User Fairness in Recommender Systems. WWW (Companion Volume) 2018: 101-102 (2018).
55. Mingming Li, Jiao Dai, Fuqing Zhu, Liangjun Zang, Songlin Hu, Jizhong Han: A Fuzzy Set Based Approach for Rating Bias. AAAI 2019: 9969-9970
(2019).
56. Daryl Lim, Julian J. McAuley, Gert R. G. Lanckriet: Top-N Recommendation with Missing Implicit Feedback. RecSys 2015: 309-312 (2015).
57. Xiao Lin, Min Zhang, Yongfeng Zhang, Zhaoquan Gu, Yiqun Liu, Shaoping Ma: Fairness-Aware Group Recommendation with Pareto-Efﬁciency.
RecSys 2017: 107-115 (2017).
58. Kun Lin, Nasim Sonboli, Bamshad Mobasher, Robin Burke: Crank up the Volume: Preference Bias Ampliﬁcation in Collaborative
Recommendation. RMSE@RecSys 2019 (2019).
59. Dugang Liu, Chen Lin, Zhilin Zhang, Yanghua Xiao, Hanghang Tong: Spiral of Silence in Recommender Systems. WSDM 2019: 222-230 (2019a).
60. Weiwen Liu, Jun Guo, Nasim Sonboli, Robin Burke, Shengyu Zhang: Personalized fairness-aware re-ranking for microlending. RecSys 2019:
467-471 (2019b).
61. Jiaqi Ma, Zhe Zhao, Xinyang Yi, Ji Yang, Minmin Chen, Jiaxi Tang, Lichan Hong, Ed H. Chi: Off-policy Learning in Two-stage Recommender
Systems. WWW 2020: 463-473 (2020).
62. Benjamin M. Marlin, Richard S. Zemel, Sam T. Roweis, Malcolm Slaney: Recommender Systems, Missing Data and Statistical Model Estimation.
IJCAI 2011: 2686-2691 (2011).

References #8
63. Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, Aram Galstyan. A survey on bias and fairness in machine learning. arXiv
preprint arXiv:1908.09635 (2019).
64. Rishabh Mehrotra, James McInerney, Hugues Bouchard, Mounia Lalmas, Fernando Diaz: Towards a Fair Marketplace: Counterfactual
Evaluation of the trade-off between Relevance, Fairness & Satisfaction in Recommendation Systems. CIKM 2018: 2243-2251 (2018).
65. Silvia Milano, Mariarosaria Taddeo, Luciano Floridi. Recommender systems and their ethical challenges. AI & Soc (2020).
66. Natwar Modani, Deepali Jain, Ujjawal Soni, Gaurav Kumar Gupta, Palak Agarwal: Fairness Aware Recommendations on Behance. PAKDD (2)
2017: 144-155 (2017).
67. Azadeh Nematzadeh, Giovanni Luca Ciampaglia, Filippo Menczer, Alessandro Flammini: How algorithmic popularity bias hinders or promotes
quality. CoRR abs/1707.00574 (2017).
68. Alexandra Olteanu, Carlos Castillo, Fernando Diaz, Emre Kiciman: Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries.
Frontiers Big Data 2: 13 (2019)
69. Zohreh Ovaisi, Ragib Ahsan, Yifan Zhang, Kathryn Vasilaky, Elena Zheleva: Correcting for Selection Bias in Learning-to-rank Systems. WWW
2020: 1863-1873 (2020).
70. Anish Anil Patankar, Joy Bose, Harshit Khanna: A Bias Aware News Recommendation System. ICSC 2019: 232-238 (2019).
71. Bruno Pradel, Nicolas Usunier, Patrick Gallinari: Ranking with non-random missing ratings: inﬂuence of popularity and positivity on evaluation
metrics. RecSys 2012: 147-154 (2012).

References #9
72. Gourab K. Patro, Arpita Biswas, Niloy Ganguly, Krishna P. Gummadi, Abhijnan Chakraborty: FairRec: Two-Sided Fairness for Personalized
Recommendations in Two-Sided Platforms. WWW 2020: 1194-1204 (2020).
73. Gourab K. Patro, Abhijnan Chakraborty, Niloy Ganguly, Krishna Gummadi. Incremental Fairness in Two-Sided Market Platforms: On Smoothly
Updating Recommendations. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 01, pp. 181-188).
74. Selwyn Piramuthu, Gaurav Kapoor, Wei Zhou, Sjouke Mauw: Input online review data and related bias in recommender systems. Decis. Support
Syst. 53(3): 418-424 (2012).
75. Guilherme Ramos and Ludovico Boratto, Reputation (In)dependence in Ranking Systems: Demographics Influence Over Output Disparities, in
Proceedings of the 43rd International ACM SIGIR Conference on Researchand Development in Information Retrieval, SIGIR 2020 (2020a)
76. Guilherme Ramos, Ludovico Boratto, Carlos Caleiro. On the negative impact of social influence in recommender systems: A study of bribery in
collaborative hybrid algorithms. Information Processing & Management, 57(2), 102058 (2020).
77. Bashir Rastegarpanah, Krishna P. Gummadi, Mark Crovella: Fighting Fire with Fire: Using Antidote Data to Improve Polarization and Fairness of
Recommender Systems. WSDM 2019: 231-239
78. Yehezkel S. Resheff, Yanai Elazar, Moni Shahar, Oren Sar Shalom: Privacy and Fairness in Recommender Systems via Adversarial Training of
User Representations. ICPRAM 2019: 476-482
79. Francesco Ricci, Lior Rokach, Bracha Shapira: Recommender Systems: Introduction and Challenges. Recommender Systems Handbook 2015:
1-34 (2015).

References #10
80. Yuta Saito, Suguru Yaginuma, Yuta Nishino, Hayato Sakata, Kazuhide Nakata: Unbiased Recommender Learning from Missing-Not-At-Random
Implicit Feedback. WSDM 2020: 501-509 (2020).
81. Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, Thorsten Joachims: Recommendations as Treatments: Debiasing
Learning and Evaluation. ICML 2016: 1670-1679 (2016).
82. Christina Schwind, Jürgen Buder: Reducing conﬁrmation bias and evaluation bias: When are preference-inconsistent recommendations
effective - and when not? Comput. Hum. Behav. 28(6): 2280-2290 (2012).
83. Dimitris Serbos, Shuyao Qi, Nikos Mamoulis, Evaggelia Pitoura, Panayiotis Tsaparas: Fairness in Package-to-Group Recommendations. WWW
2017: 371-379 (2017).
84. Ashudeep Singh, Thorsten Joachims: Policy Learning for Fairness in Ranking. NeurIPS 2019: 5427-5437 (2019).
85. Ashudeep Singh, Thorsten Joachims: Fairness of Exposure in Rankings. KDD 2018: 2219-2228 (2018).
86. Nasim Sonboli, Robin Burke: Localized Fairness in Recommender Systems. UMAP (Adjunct Publication) 2019: 295-300 (2019).
87. Harald Steck: Item popularity and recommendation accuracy. RecSys 2011: 125-132 (2011).
88. Maria Stratigi, Haridimos Kondylakis, Kostas Stefanidis: Fairness in Group Recommendations in the Health Domain. ICDE 2017: 1481-1488
(2017).
89. Erich Teppan, Marcus Zanker, M: Decision biases in recommender systems. Journal of Internet Commerce, 14(2), 255-275 (2015).
90. Mucun Tian, Michael D. Ekstrand: Estimating Error and Bias in Ofﬂine Evaluation Results. CHIIR 2020: 392-396 (2020).

References #11
91. Songül Tolan: Fair and Unbiased Algorithmic Decision Making: Current State and Future Challenges. CoRR abs/1901.04730 (2019).
92. Raciel Yera Toledo, Yailé Caballero Mota, Luis Martínez-López: Correcting noisy ratings in collaborative recommender systems. Knowl. Based
Syst. 76: 96-108 (2015).
93. Andreu Vall, Massimo Quadrana, Markus Schedl, Gerhard Widmer: Order, context and popularity bias in next-song recommendations. Int. J.
Multim. Inf. Retr. 8(2): 101-113 (2019).
94. Mengting Wan, Jianmo Ni, Rishabh Misra, Julian J. McAuley: Addressing Marketing Bias in Product Recommendations. WSDM 2020: 618-626
(2020).
95. Xiaojie Wang, Rui Zhang, Yu Sun, Jianzhong Qi: Doubly Robust Joint Learning for Recommendation on Data Missing Not at Random. ICML
2019: 6638-6647 (2019).
96. Jacek Wasilewski, Neil Hurley: Are You Reaching Your Audience?: Exploring Item Exposure over Consumer Segments in Recommender
Systems. UMAP 2018: 213-217 (2018).
97. Bin Xia, Junjie Yin, Jian Xu, Yun Li: WE-Rec: A fairness-aware reciprocal recommendation based on Walrasian equilibrium. Knowl. Based Syst.
182 (2019).
98. Yuanbo Xu, Yongjian Yang, Jiayu Han, En Wang, Fuzhen Zhuang, Hui Xiong: Exploiting the Sentimental Bias between Ratings and Reviews for
Enhancing Recommendation. ICDM 2018: 1356-1361 (2018).
99. Himank Yadav, Zhengxiao Du, Thorsten Joachims: Fair Learning-to-Rank from Implicit Feedback. CoRR abs/1911.08054 (2019).

References #12
100. Longqi Yang, Yin Cui, Yuan Xuan, Chenyang Wang, Serge J. Belongie, Deborah Estrin: Unbiased offline recommender evaluation for
missing-not-at-random implicit feedback. RecSys 2018: 279-287 (2018).
101. Ke Yang, Julia Stoyanovich: Measuring Fairness in Ranked Outputs. SSDBM 2017: 22:1-22:6 (2017).
102. Sirui Yao, Bert Huang: Beyond Parity: Fairness Objectives for Collaborative Filtering. NIPS 2017: 2921-2930 (2017).
103. Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, Ed H. Chi: Sampling-bias-corrected
neural modeling for large corpus item recommendations. RecSys 2019: 269-277 (2019).
104. Meike Zehlike, Carlos Castillo: Reducing Disparate Exposure in Ranking: A Learning To Rank Approach. WWW 2020: 2849-2855 (2020).
105. Meike Zehlike, Francesco Bonchi, Carlos Castillo, Sara Hajian, Mohamed Megahed, Ricardo Baeza-Yates: FA*IR: A Fair Top-k Ranking Algorithm.
CIKM 2017: 1569-1578 (2017).
106. Shuai Zhang, Lina Yao, Aixin Sun, Yi Tay: Deep Learning Based Recommender System: A Survey and New Perspectives. ACM Comput. Surv.
52(1): 5:1-5:38 (2019)
107. Xiaoying Zhang, Hong Xie, Junzhou Zhao, John C. S. Lui: Modeling the Assimilation-Contrast Effects in Online Product Rating Systems:
Debiasing and Recommendations. IJCAI 2018: 5409-5413 (2018).
108. Xing Zhao, Ziwei Zhu, Majid Alfifi, James Caverlee: Addressing the Target Customer Distortion Problem in Recommender Systems. WWW
2020: 2969-2975 (2020).
109. Yong Zheng, Tanaya Dave, Neha Mishra, Harshit Kumar: Fairness In Reciprocal Recommendations: A Speed-Dating Study. UMAP (Adjunct
Publication) 2018: 29-34 (2018).

References #13
110. Ziwei Zhu, Xia Hu, James Caverlee: Fairness-Aware Tensor-Based Recommendation. CIKM 2018: 1153-1162 (2018).

Tutorial on Bias in Rec Sys @ UMAP2020

More Related Content

What's hot (20)

Similar to Tutorial on Bias in Rec Sys @ UMAP2020 (20)

Recently uploaded (20)

Tutorial on Bias in Rec Sys @ UMAP2020