Interactive Clustering Barna Saha Clustering Learning over Noisy - - PowerPoint PPT Presentation

▶

Jun 23, 2023 183 likes •323 views

Interactive Clustering Barna Saha Clustering Learning over Noisy Data Learn a classifier or find clusters over noisy/uncertain data Noise comes from using similarity func5onsadd an edge between two images if they represent the same

SLIDE 1

Interactive Clustering

Barna Saha

SLIDE 2

Clustering

SLIDE 3

Learning over Noisy Data

Noise comes from using similarity func5ons—add an edge between two images if they represent the same monument—clusters could be erroneous Learn a classifier or find clusters over noisy/uncertain data

SLIDE 4

Learn a classifier or do clustering over noisy/uncertain data

Noise comes from inherent data errors/missing a?ributes—clustering collabora5on network

btained from DBLP could be erroneous.

Learning over Noisy Data

Learn a classifier or find clusters over noisy/uncertain data

SLIDE 5

Further Applications

Linking Census Records
Public Health
Web search
Comparison shopping
Spam Detec5on
Machine Reading
IP Aliasing
……..

SLIDE 6

Query complexity of optimal strategy?

SLIDE 7

Query complexity of optimal strategy?

Davidson, Khanna, Milo, Roy, 2014

SLIDE 8

Faulty Oracle

SLIDE 9

Faulty Oracle

Repeat the same ques5on. Assuming p=q, repeat each ques5on (say) 24log n/(1-2p)2 5mes

SLIDE 10

Faulty Oracle

SLIDE 11

Faulty Oracle: No Resampling

Find seed nodes for each cluster
If we can find 24 log n/(1-2p)2 seed nodes from each cluster

then we are done! [Why?]

….. ………………..

SLIDE 12

Faulty Oracle: How to Cind seed nodes?

Let N=O(k2log n/(1-2p)4)
Select N nodes and ask all possible pairwise queries among

these nodes.

Run correla5on clustering algorithm in this small set of nodes
Each cluster returned by the correla5on clustering that has

size at least 24 log n/(1-2p)2 act as a seed

SLIDE 13

Faulty Oracle: How to Cind seed nodes?

Let N=O(k2log n/(1-2p)4)
Select N nodes and ask all possible pairwise queries among

these nodes.

Run correla5on clustering algorithm in this small set of nodes
Each cluster returned by the correla5on clustering that has

size at least 24 log n/(1-2p)2 act as a seed

Some intui5on on the analysis: If we know all the query results, correla5on clustering gives the maximum likelihood es5mator. Moreover, it is an instance of correla5on clustering where errors are random— we know how to solve it!