BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1037
DTSTAMP:20230914T125948Z
SUMMARY:PIDForest: Anomaly detection via Partial Identification
DESCRIPTION:Speaker: Parikshit Gopalan (VMware Research\nPalo Alto\, CA\, U
 .S.)\n\nAbstract: \nAbstract: Anomaly detection is a ubiquitous problem in
  machine learning. Here one is given a large population of points\, we may
  not have much knowledge about their structure a priori. The goal is to fi
 gure out what is "typical" of the datatset\, and what points are atypical 
 or anomalous.  This talk will explore a framework called Partial Identifi
 cation for identifying anomalies in a large datatset.\n\nWe propose a defi
 nition for “anomalousness” that captures the intuition that anomalies 
 are easy to distinguish from the overwhelming majority of points by relati
 vely few attribute values: we call this partial identification. Our notio
 n is inspired by the notion of Partial IDs that were studied by Yehudayoff
  and Wigderson in the context of population recovery. Formalizing this int
 uition\, we propose a geometric anomaly measure for a point that we call P
 IDScore\, which measures for the minimum density of data points over all s
 ubcubes containing the point. We present PIDForest: a random forest based
  algorithm that finds anomalies based on this definition and show that it 
 performs favorably in comparison to several popular anomaly detection meth
 ods\, across a broad range of benchmarks.\n\nBased on joint work with Udi 
 Wieder (VMware) and Vatsal Sharan (Stanford) that appeared in NeurIPS 2019
 .\n
URL:https://www.tcs.tifr.res.in/web/events/1037
DTSTART;TZID=Asia/Kolkata:20200113T143000
DTEND;TZID=Asia/Kolkata:20200113T153000
LOCATION:A-201 (STCS Seminar Room)
END:VEVENT
END:VCALENDAR