Online k-means Clustering on Arbitrary Data Streams - Lecture
Offered By: USC Probability and Statistics Seminar via YouTube
Course Description
Overview
Explore online k-means clustering for arbitrary data streams in this 37-minute USC Probability and Statistics Seminar talk by Robi Bhattacharjee. Delve into a novel approach for achieving clustering loss that is comparable to the best possible loss using k fixed points in hindsight. Learn about a proposed data parameter, Λ(X), and its implications for algorithm performance. Discover a randomized algorithm that achieves O(Λ(X)+L(X,OPTk)) clustering loss while maintaining O(kpoly(logn)) memory and cluster centers. Understand how this algorithm achieves polynomial space and time complexity without making assumptions on input data. Follow the presentation through key concepts including the online setting, lower bounds, streaming challenges, and innovative ideas for center management and point deletion. Gain insights into future directions for research in this field.
Syllabus
Intro
k-means clustering
The online setting
The Goal(s) Cohen-Addad et. al. 2021
A troubling example
Lower Bound
A natural starting point: streaming
A difficult case for streaming
Idea 1: don't remove centers
Proof Sketch
We still have problems on pathological examples.
Idea 2: Using the scale to delete points.
The Lemma Revisited
Our Algorithm's Performance
Proof idea
Future Directions
Taught by
USC Probability and Statistics Seminar
Related Courses
Divide and Conquer, Sorting and Searching, and Randomized AlgorithmsStanford University via Coursera Unpredictable? Randomness, Chance and Free Will
National University of Singapore via Coursera Biology Meets Programming: Bioinformatics for Beginners
University of California, San Diego via Coursera Finding Hidden Messages in DNA (Bioinformatics I)
University of California, San Diego via Coursera Algorithms for Big Data
Indian Institute of Technology Madras via Swayam