STA663-Final-Project: Scalable K-Means

Team Members: Yunlu Hao, Lianghui Li

Include:

• Kmeans.ipynb: main file
• data: include all data files
• test: intermediate code
• readme.md

Package:

• package_kmeans

This is the final project for course STA663. In this project, we implement the K-means++ and K-means||algorithm in Python following Bahmani's paper "Scalable k-means++" and speed up using Cython,JIT and multi-core processing.

We apply it to the simulated GAUSSMIXTURE dataset, SPAM and Housing dataset from UC Irvine Machine Learning repository. In this part, we compare clustering cost and runtime of the k-means|| algorithm with these of random initialization and the k-means++.

From the implementation on the simulated dataset and the real-world dataset, the k-means|| and k-means++ find a better initial centroids than random in most cases, which leads to a better final clustering performance. However, for our algorithm, k-means++ runs faster than k-means|| in most cases, in future anlysis, we should focus more on how to implement k-means|| in a more robust setting(in terms of cost reduction and better parallel application.

We also created an easily instable package for our functions, which can be found in https://github.com/yumegaaru/package_kmeans

To install, please go to Terminal and type in:

!pip install --upgrade pip git+git://github.com/yumegaaru/kmeans_package.git

then run the following code in your python:

import kemeans_package

Data and test code used in our analysis are also provided in folders under this repository.