The Mean Shift is a powerful unsupervised learning algorithm that's used to cluster datapoints. It considers the distribution of datapoints as a probability-density function and tries to find the modes in the feature space. These modes are basically points corresponding to local maxima. The main advantage of Mean Shift algorithm is that we are not required to know the number of clusters beforehand.
Let's say that we have a set of input points, and we are trying to find clusters in them without knowing how many clusters we are looking for. Mean Shift algorithm considers these points to be sampled from a probability density function. If there are clusters in the datapoints, then they correspond to the peaks of that probability-density function. The algorithm starts from random points and iteratively converges toward these peaks. You can learn more about it at http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/TUZEL1/MeanShift.pdf.
- The full code for this recipe is given in the
mean_shift.pyfile that's already provided to you. Let's look at how it's built. Create a new Python file, and import a couple of required packages:import numpy as np from sklearn.cluster import MeanShift, estimate_bandwidth import utilities
- Let's load the input data from the
data_multivar.txtfile that's already provided to you:# Load data from input file X = utilities.load_data('data_multivar.txt') - Build a Mean Shift clustering model by specifying the input parameters:
# Estimating the bandwidth bandwidth = estimate_bandwidth(X, quantile=0.1, n_samples=len(X)) # Compute clustering with MeanShift meanshift_estimator = MeanShift(bandwidth=bandwidth, bin_seeding=True)
- Train the model:
meanshift_estimator.fit(X)
- Extract the labels:
labels = meanshift_estimator.labels_
- Extract the centroids of the clusters from the model and print out the number of clusters:
centroids = meanshift_estimator.cluster_centers_ num_clusters = len(np.unique(labels)) print "Number of clusters in input data =", num_clusters
- Let's go ahead and visualize it:
# Plot the points and centroids import matplotlib.pyplot as plt from itertools import cycle plt.figure() # specify marker shapes for different clusters markers = '.*xv'
- Iterate through the datapoints and plot them:
for i, marker in zip(range(num_clusters), markers): # plot the points belong to the current cluster plt.scatter(X[labels==i, 0], X[labels==i, 1], marker=marker, color='k') # plot the centroid of the current cluster centroid = centroids[i] plt.plot(centroid[0], centroid[1], marker='o', markerfacecolor='k', markeredgecolor='k', markersize=15) plt.title('Clusters and their centroids') plt.show() - If you run this code, you will get the following output:
