homework4sol

# homework4sol - CS573: Homework 4 Solution Part of the...

This preview shows pages 1–4. Sign up to view the full content.

CS573: Homework 4 Solution Part of the solution is based on Wahbeh Q.’s submission 1 Between cluster distances (4 pts) Let cluster C i contain n i samples, and let d ij be some measure of distance between two clusters C i and C j . In general, one might expect that if C i and C j are merged to form a new cluster C k , then the distance from C k to some other cluster C h is not simply related to d hi and d hj . However, consider the equation: d hk = α i d hi + α j d hj + βd ij + γ | d hi - d hj | Show that the following choices for the coeﬃcients α i j ,β,γ lead to the distance functions indi- cated. 1. Single-link: α i = α j = 0 . 5 = 0 = - 0 . 5 Sample solution: d hk = α i d hi + α j d hj + ij + γ | d hi - d hj | = 0 . 5 d hi + 0 . 5 d hj - 0 . 5 | d hi - d hj | = ( 0 . 5 d hi + 0 . 5 d hj - 0 . 5 d hi + 0 . 5 d hj , if d hi d hj 0 . 5 d hi + 0 . 5 d hj - 0 . 5 d hj + 0 . 5 d hi , if d hi < d hj = ( d hj , if d hi d hj d hi , if d hi < d hj = min { d hj ,d hi } = min { min u C h ,v C j d ( u,v ) , min u C h ,v C i d ( ) } = min u C h ,v C j C i { d ( ) } = min u C h ,v C k { d ( ) } = d hk 2. Complete-link: α i = α j = 0 . 5 = 0 = +0 . 5 1

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Sample solution: d hk = α i d hi + α j d hj + βd ij + γ | d hi - d hj | = 0 . 5 d hi + 0 . 5 d hj + 0 . 5 | d hi - d hj | = ( 0 . 5 d hi + 0 . 5 d hj + 0 . 5 d hi - 0 . 5 d hj , if d hi d hj 0 . 5 d hi + 0 . 5 d hj + 0 . 5 d hj - 0 . 5 d hi , if d hi < d hj = ( d hi , if d hi d hj d hj , if d hi < d hj = max { d hj ,d hi } = max { max u C h ,v C j d ( u,v ) , max u C h ,v C i d ( ) } = max u C h ,v C j C i { d ( ) } = max u C h ,v C k { d ( ) } = d hk 3. Average-link: α i = n i n i + n j j = n j n i + n j = γ = 0 Sample solution: d hk = α i d hi + α j d hj + ij + γ | d hi - d hj | = n i n i + n j d hi + n j n i + n j d hj = n i n i + n j ± u C h ,v C i d ( ) n i ² + n j n i + n j u C h ,v C j d ( ) n j ! = u C h ,v C i C j d ( ) n i + n j = u C h ,v C k d ( ) n k = d hk 4. Between-cluster distance (i.e., squared Euclidean distance between centroids): α i = n i n i + n j j = n j n i + n j = - α i α j = 0 Sample solution: Let m i denote the centroid of C i , which is 1 n i u C i u . Furthermore, let d 2 Eu denote the squared Euclidean distance 2
d hk = α i d hi + α j d hj + βd ij + γ | d hi - d hj | = n i n i + n j d hi + n j n i + n j d hj - n i n j ( n i + n j ) 2 d ij = n i d 2 Eu ( m h ,m i ) n i + n j + n j d 2 Eu ( m h j )) n i + n j - n i n j d 2 Eu ( m i j ) ( n i + n j ) 2 = n i ( m h - m i ) 2 n i + n j + n j ( m h - m j ) 2 n i + n j - n i n j ( m i - m j ) 2 ( n i + n j ) 2 = n i m 2 h - 2 n i m h m i + n i m 2 i + n j m 2 h - 2 n j m h m j + n j m 2 j n i + n j - n i n j ( m i - m j ) 2 ( n i + n j ) 2 = m 2 h ( n i + n j ) + n i m 2 i + n j m 2 j - 2 m h ± u C i u + u C j u ² n i + n j - n i n j ( m i - m j ) 2 ( n i + n j ) 2 = m 2 h ( n i + n j ) + n i m 2 i + n j m 2 j - 2 m h ± u C i C j u ² n i + n j - n i n j ( m i - m j ) 2 ( n i + n j ) 2 = m 2 h - 2 m h m k + n i m 2 i + n j m 2 j n i + n j - n i n j ( m i - m j ) 2 ( n i + n j ) 2 = m 2 h - 2 m h m k + ( n i + n j ) n i m 2 i + ( n i + n j ) n j m 2 j - n i n j ( m i - m j ) 2 ( n i + n j ) 2 = m 2 h - 2 m h m k + n 2 i m 2 i + 2 n i n j m i m j + n 2 j m 2 j ( n i + n j ) 2 = m 2 h - 2 m h m k + ( n i m i + n j m j ) 2 ( n i + n j ) 2 = m 2 h - 2 m h m k + u C i C j u n i + n j !

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
This is the end of the preview. Sign up to access the rest of the document.

## This note was uploaded on 03/13/2012 for the course CS 573 taught by Professor Staff during the Fall '08 term at Purdue.

### Page1 / 11

homework4sol - CS573: Homework 4 Solution Part of the...

This preview shows document pages 1 - 4. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online