Anant00 / K-Means-Clustering-From-Scratch

K-Means

K-Means-Clustering-From-Scratch

Data Mining: Using K-Means clustering to gain inisghts on an Airbnb dataset from Kaggle

Background

This K-Means algorithm is written specifically for an input pandas DataFrame.

Notes about K-Means:

Pros

a. Simple, easy to implement
b. Efficient
c. Often terminates at local optimum
d. Works well in practice
e. If we start frommg good initial point we may converge faster

Cons

a. Applicable to only objects in continuous n-dim space
b. Does not work with categorical data (unless transformed)
c. Need to specify K in advance
d. Sensitive to noisy data and outliers
e. Good when clusters are Gaussian
f. Each time we run algorithm, we may get different answer since we start at random k
g. You might have an empty cluster due to bad starting point

Algorithm:

Initialize cluster centroids:
a. Pick k points at random and set as centroids represented each cluster
Repeat while cluster assignments don't change
b. Assign each point to nearest centroid b. Given new assignments, compute new cluster centroids as mean of all possible
points in cluster

1. What neighborhoods are the most expensive?
2. What room_types are most desirable?
3. Cluster listing based on room_type / price / num_reviews (popularity)

Getting Started

1. Install Python 3

brew install python3

Running the Code

The KMeansAlgorithm.py will run the algorithm using this Airbnb dataset.

The command line arguments require (1) K clusters (int) and (2) number of iterations (int). A good number of iterations is 100 for the stopping criterion. When the centroids remain the same after iterations, this signifies convergence. Below is and example of running the algorithm with 4 clusters and 100 iterations.

In command line, run: python KMeansAlgorithm.py 4 100

Sample Output:

K-Means Plot with K=2-8 clusters:

Elbow Plot with K=4 clusters: