Tutorial on
Bias in Personalized Rankings: Concepts to Code

to be held as part of the 20th IEEE International Conference on Data Mining (ICDM2020)

November, 2020 - ONLINE

Introduction

This tutorial aims to provide the ICDM community with a comprehensive survey on data and algorithmic bias in personalized rankings, with a focus on state-of-the-art recommendation algorithms and systems. First, we present conceptual fundamentals on bias based on the existing literature, including a broad discussion on real-world perspectives, ethical aspects, and system's objectives impacted by this phenomenon. Then, we dive deeper into concepts, techniques, metrics, and frameworks that allow to detect, understand, and mitigate bias along the algorithm design pipeline, by presenting the theoretical foundations behind state-of-the-art approaches and their implementations. We showcase hands-on examples of pre-, in-, and post-processing bias mitigation techniques by means of open-source tools and public datasets, engaging tutorial participants in mitigation design and in articulating impacts on stakeholders. Finally, we shed light on how the research in this novel and fast-growing field has opened challenges and opportunities for the ICDM community.

Target Audience

This tutorial is accessible to researchers, industry technologists and practitioners. For people not familiar with rankings, this tutorial covers necessary background material. No prior knowledge on bias is assumed.

One aspect relevant from the outline is that bias is a highly interdisciplinary topic, touching on several dimensions beyond algorithms. Hence, our tutorial is of interest for an interdisciplinary audience attending ICDM, with different backgrounds.

After this tutorial, attendees will be able to understand key aspects of bias in personalized rankings, materialize biases into underlying systems, play with mitigation and articulate impacts on stakeholders, identify challenges and opportunities.

Outline

Due to the ongoing worldwide COVID-19 situation, ICDM 2020 will take place online.

Timing Content
5 mins Welcome and Presenters' Introduction
130 mins Session I: Foundations
Recommendation Principles
  • Recommendation principles. To introduce the problems associated to algorithmic bias, we will present the recommendation task as the generation of the most effective personalized ranking for a user, as in modern recommender systems.
  • Multi-sided recommendation aspects. Recommender systems have an impact on multiple actors, namely consumers, providers, system's owners. We will present these actors and the phases of the recommendation process where they have a role (design, algorithm, and evaluation).
Algorithmic Bias Foundations
  • Motivating examples. We will present real-world examples where bias can impact recommendation, considering domains such as music, education, social platforms, and recruiting.
  • Perspectives impacted by bias. Bias has an impact on several perspectives such as the economy, law, society, security, technology, and psychology.
  • Ethical aspects influenced by bias. Bias can have an impact at the ethical level and lead to issues such as recommendation of inappropriate content, lack of privacy, violation of autonomy and identity, introduction of opacity, lack of fairness, or the compromising of users' social relationships.
  • Objectives influenced by bias. We will present recommendation objectives influenced by bias (utility, coverage, diversity, novelty, visibility, exposure) and provide examples of related work.
Bias through the Pipeline
  • Recommendation pipeline. We will provide an initial overview of the recommendation pipeline, to characterize how bias can exist at several stages, namely, data acquisition and storage, data preparation, model training, model prediction, model evaluation, and recommendation delivery.
  • Types of bias associated to the pipeline. We explore the types of bias that can emerge at different stages of the pipeline, i.e., those associated to the users, platforms, data collection, data preparation, model exploitation, and model evaluation.
Discrimination
  • Types of discrimination. When bias affects users' sensitive attributes, it may lead to discrimination. We present concepts, such as direct/indirect discrimination, its granularity (group, individual, and subgroup discrimination), types of disparity (disparate treatment, impact, and mistreatment).
  • Definitions of fairness. We will present definitions of fairness and different classes in which thy can be categorized (equalized odds, equalized opportunity, demographic parity, fairness through (un)awareness, equality of treatment).
Mitigation Design
  • Bias-aware process pipeline. Intervention strategies to mitigate algorithmic bias require an analysis of where and how bias might affect the system. We present a pipeline to support mitigation design.
  • Techniques for bias treatment. We will present the three main classes of mitigation techniques (pre-, in-, and post-processing), along with examples of solutions proposed for recommender systems.
  • Real-world applications. We will present examples of real-world platforms, such as LinkedIn and Spotify, and of their approaches to deal with bias.
Questions and Discussion
30 mins Coffee Break
110 mins Session II: Hands-on Studies
Recommender Systems in Practice
  • Data preparation starting from public datasets (i.e., COCO and Movielens datasets).
  • Model definition (e.g., user/item embeddings, layers stacking) and training (e.g., epochs, loss, optimizer)
  • User-item relevance matrix computation from a pre-trained model (e.g., model load, predictions).
  • Model evaluation oriented to utility (e.g., NDCG, beyond-accuracy metrics).
Investigation on Item Popularity Bias
  • Definition and characterization of item popularity biases in interactions and recommendations.
  • Application of mitigation techniques based on pre-, in-, and post-processing.
  • Comparison of mitigation techniques based on bias and recommendation utility trade-offs.
  • Comparison of mitigation techniques on beyond-utility metrics (e.g., coverage, diversity, novelty).
Investigation on Item Provider Fairness
  • Association of items to providers and sensitive attributes.
  • Characterization of providers representation in the catalog and in the interactions.
  • Identification of minority providers, both at individual and group level.
  • Definition and measurement of item provider unfairness on recommendations.
  • Application of mitigation techniques based on pre-, in-, and post-processing.
  • Comparison of mitigation techniques based on fairness and recommendation utility trade-offs.
  • Comparison of mitigation techniques on beyond-utility metrics (e.g., coverage, diversity, novelty).
10 mins Open Issues and Research Challenges
15 mins Questions and Discussion

Material

Lecture slides, Github repository, and Jupyter notebooks will be made available to attendees before the tutorial.

Presenters

Ludovico Boratto

Ludovico Boratto
EURECAT - Centre Tecnòlogic de Catalunya (Spain)

Ludovico Boratto is senior research scientist in at EURECAT. His research focuses on recommender systems and on their impact on stakeholders. His research has been published in top-tier conferences and journals. He is editor of the book "Group Recommender Systems: An Introduction" (Springer). He is editorial board member of the "Information Processing & Management" journal (Elsevier) and guest editor of other special issues. He is regularly PC member of the main Data Mining conferences. In 2012, he got a Ph.D. at the University of Cagliari, where he was research assistant until May 2016.



Mirko Marras

Mirko Marras
University of Cagliari (Italy)

Mirko Marras is a Postdoctoral Researcher at the University of Cagliari (Italy). His research focuses on machine learning for recommender systems, with particular attention to bias issues. He authored papers in top-tier journals, such as Pattern Recognition Letters and Computers Human Behavior. He gave talks and demos at international conferences and workshops, e.g., TheWebConf2018, ECIR2019, and INTERSPEECH2019. He is PC member of major conferences, e.g., ACL, AIED, EDM, ECML-PKDD, EMNLP, ITICSE, ICALT, UMAP. He co-chaired the BIAS2020 workshop at ECIR2020 and gave a tutorial on Bias in Recommender Systems at UMAP2020. In 2020, he received a Doctoral Degree from University of Cagliari.

Contacts

Please, reaching out to us at ludovico.boratto@acm.org and mirko.marras@unica.it.

Past Editions

We also invite you to check out previous editions of our similar tutorials: