Seminars

View all Seminars  |  Download ICal for this event

Eigenvectors Made Many People Rich, but Can They be Fair to Each Individual? On Constrained Spectral Clustering with Provable Guarantees

Series: CSA Faculty Colloquium

Speaker: Prof. Ambedkar Dukkipati, Professor, Dept. of C.S.A, IISc

Date/Time: Jul 05 16:00:00

Location: CSA Lecture Hall (Room No. 112, Ground Floor)

Abstract:
-Objectivity- comes to our mind when we replace humans with intelligent machines in decision-making. However, recently, it has been shown that machine learning algorithms can lead to -unfair- decisions. For example, it may be desirable to ensure that clusters discovered by an algorithm do not have a gender bias. What about being fair and representing each individual?
<br>
Spectral graph methods for detecting communities in networks have been very successful. Statistical analysis of spectral algorithms involves assuming that the data or network is generated from a random model with a planted partition and then deriving bounds on the error obtained by a particular spectral algorithm. In this talk, I will address the following problem. Can the spectral methods provide consistent solutions for finding clusters representing each individual where sensitive attributes indirectly manifest in an auxiliary representation graph? I will present constrained spectral clustering algorithms for this task and show that these algorithms are weakly consistent under a -fair- planted partition model induced by the representation graph.
<br>
I will also provide the necessary background at the beginning of the talk by briefly introducing spectral graph theory results that lead to an approximate method for graph partitioning.

Speaker Bio:
Ambedkars research lies in Statistical Machine Learning, Reinforcement Learning, Network Analysis, and Predictive Analytics. He is interested in developing algorithms for practical AI problems with theoretical guarantees. He holds a B.Tech in Computer Science from Indian Institute of Technology, Madras, and a Masters and Ph.D in Computer Science from Indian Institute of Science. After his postdoctoral research at EURANDOM in the Netherlands, he joined as a faculty at CSA, IISc.

Host Faculty: Prof. Arindam Khan