This document discusses future directions for deep recommender systems across five key dimensions: privacy, explainability, fairness, robustness, and accountability. It notes that recommender systems rely heavily on private user data and explores how to protect privacy while leveraging data advantages. Explainability is important to address issues around transparency, trust, and debugging models. Fairness aims to avoid discrimination and ensure equitable decision making. Robustness examines adversarial attacks and fake user profiles. Accountability determines responsibility across system designers, decision makers, deployers, and auditors. Interactions among these dimensions must also be considered to guide future work in developing more trustworthy recommender systems.
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
Trustworthy Recommender Systems
1. Deep Recommendations:
Future Directions
1
Data Science and Engineering Lab
Tutorial website: https://advanced-recommender-systems.github.io/ijcai2021-tutorial/
Wenqi Fan
The Hong Kong Polytechnic University
https://wenqifan03.github.io, wenqifan@polyu.edu.hk
4. Privacy Issue
4
q The success of recommender systems
heavily relies on data that might
contain private and sensitive
information.
q Can we still take the advantages of
data while effectively protecting the
privacy?
6. Black-box AI creates confusion and doubt
Explainability Issue
6
User with
a Task
IT & Operations
Can I trust our
system’s decisions?
Business Owner
Data Scientists
Are these system
decisions fair?
Internal Audit, Regulators
Customer Support
How do I answer this
customer complaint?
How do I monitor and
debug this model?
Is this the best model
that can be built?
Output
The Need for Explainable Recommendation
From Black-box
to “Transparent”
Yongfeng Zhang, et.al, Explainable Recommendation: A Survey and New Perspectives, 2020.
8. Discrimination & Fairness Issue
8
Job recommendation
(Lambrecht et al., 2019)
Lambrecht, et al. "Algorithmic bias? An empirical study of apparent gender-based discrimination in the display of STEM career ads." 2019.
Bias and Debias in Recommender System: A Survey and Future Directions, 2021.
9. § A recommender system should avoid discriminatory behaviors in
human-machine interaction.
§ A recommender system should ensure fairness in decision-making.
Non-discrimination & Fairness
11. § Accountability
Auditability & Accountability
11
Accountability: A clear responsibility distribution, which focuses on who
should take the responsibility for what impact of recommender systems.
Who is to
blame?
Violent movie
12. § Five roles in Recommender Systems
Auditability & Accountability
12
System
Designers
Decision
Makers
System
Deployers
System
Auditors
End Users
It is necessary to determine the roles and the corresponding responsibility of
different parties in the function of a recommender system.
13. Interactions Among Different Dimensions
13
How do these five dimensions influence each other?
There exist both accordance and the conflicts among the five dimensions.
14. A Survey on The Computational Perspective
14
https://arxiv.org/abs/2107.06641
Tutorial website: https://sites.google.com/msu.edu/trustworthy-ai/home
15. 15
Acknowledgements
Thanks for this great opportunity offered by IJCAI 2021 and our funding support
from MSU, AERA, NSF, ARO, DARPA,NSFC, GRF, and industrial collaborators.