Explanation and Fairness in Unsupervised
Learning
Overview: Unsupervised
learning approaches such as clustering are extensively used in AI with
explanation and fairness emerging as important challenges. We cover explainable
and fair unsupervised learning from multiple perspectives such as the
philosophical under-pinnings, algorithmic details and application areas.
Bio: Ian Davidson: I am a
Professor in the department of computer science at the University of
California, Davis since 2009 (and before that an assistant and associated
Professor). I teach AI, ML and for the past decade a course on ethics and
AI/Technology. My research is a mix of fundamental algorithmic contributions to
unsupervised learning (clustering and outlier detection) and supervised
learning (transfer, active learning) and more practical applications (with
others) in neuroscience and intelligent tutoring. With regard to fairness and
explanation in unsupervised learning, I have published papers at NIPS 2018,
IJCAI 2018, AAAI 2020 and ECAI 2020 on the topic and I have grants from NSF on
core research on the topic and NIH to apply XAI to precision medicine.
My website is here my google
scholar page is here.
Slides To Be Posted
Tutorial Outline and References
The tutorial will consist of three parts: i)
Legal, Philosophical and Ethical Overviews and Motivations, ii) Algorithmic
Details and iii) Applications and Future Work.
Motivations (15 minutes)
– Legal (in California, USA and Europe),
ethical and philosophical definitions of fairness
and explanations.
– Why is fairness and explanation so
necessary for unsupervised learning.
– The notion of protected state variables
(PSV). Examples of individual level and group level unfairness
–
Connection to
Utilitarianism and Deontological schools of normative ethics.
–
The need for explanations to algorithm designer, domain scientists and
public.
Algorithmic Details (60 minutes)
– The two approaches for adding fairness
and explanation to algorithms - (apriori) fairness/explanation by-design and
(aposteri) post-processing to increase fairness/explanation.
– Fairness in Clustering
∗ Fairness - Seminal paper encoding disparate impact in clustering [6] for
classic centroid based methods.
∗ Extensions for efficiency [1] to allow multi-state [19], multiple [2]
PSVs.
∗ Fairness methods for graph based [16] and other clustering algorithms [5].
∗ Emerging work on individual level fairness [21] [17].
∗ Algorithm independent methods of fairness - post-processing to ensure fairness
[12].
– Explanation in
Clustering
∗ Explanation -
Classic Work (Conceptual Clustering) [13] [18].
∗ Explanation using feature spaces [4].
∗ Explanation using alternative feature
sets [10], [20], [8].
– Fairness and Explanation in
Emerging Unsupervised Learning Settings.
∗ Outlier detection [7] [11].
∗ Nearest neighbor search [15]. ∗ Embedding [3].
– The Connections Between
Fairness and Explanation.
¥ Future Directions and Applications
(15 minutes)
– Applications and Transitions into Industry
∗ IBMÕs AI Fairness 360 toolkit
∗ AccentureÕs Fairness verification
toolkit
–
Open Research Questions