K-Means

Parallel k-means clustering algorithm.

Goal

General purpose k-means clustering, supporting both sequential and parallel execution (depending on input sequences) and clusters homogeneity measures.

Sample clients

Client can be run with either

• Scala, using scala target/scala-2.12/kmeans.jar [OPTIONS], or
• Java, using java -cp target/scala-2.12/kmeans.jar Client [OPTIONS]

Build a jar file:

$ ./sbt assembly

Client options:

$ scala target/scala-2.12/kmeans.jar -h

Image clustering

Original photo

K-means clustering with 5 runs, 32 clusters and default k-means++ sampling:

scala target/scala-2.12/kmeans.jar -k 32 -r 5 data/photo.jpg > photo32.png

K-means clustering with 5 runs, 16 clusters and default k-means++ sampling:

scala target/scala-2.12/kmeans.jar -k 16 -r 5 data/photo.jpg > photo16.png

K-means clustering with 5 runs, 8 clusters and default k-means++ sampling:

scala target/scala-2.12/kmeans.jar -k 8 -r 5 data/photo.jpg > photo8.png

Text clustering

Raw articles were downloaded from BBC Datasets website and pre-processed using Python's NLTK library. There are 2225 articles from five categories (business: 510, entertainment: 386, politics: 417, sport: 511, tech: 401)

K-means clustering with 20 runs, 5 clusters and default k-means++ sampling:

scala target/scala-2.12/kmeans.jar -k 5 -r 20 data/articles.txt

Cluster 0
Articles #:    420
Top 10 words:  mr, labour, election, blair, party, tory, government, brown, minister, tax

Cluster 1
Articles #:    521
Top 10 words:  game, england, play, win, match, player, cup, club, champion, injury

Cluster 2
Articles #:    391
Top 10 words:  mobile, phone, technology, user, use, computer, game, software, digital, people

Cluster 3
Articles #:    502
Top 10 words:  bank, company, growth, firm, economy, market, share, price, rise, sale

Cluster 4
Articles #:    391
Top 10 words:  film, award, star, best, band, oscar, actor, show, album, music