Learn NANI in 60 seconds!

How to use NANI in 60 seconds? Say no more!

The main idea is to use the NANI to optimize initial centroids so k-means is 100% deterministic, converges faster, and finds better solutions. Here is a simple example to get started.

The pwd of this script is $PATH/MDANCE/examples.

Let’s start with importing necessary libraries.

from matplotlib import pyplot as plt
from sklearn.cluster import KMeans
from sklearn.datasets import make_blobs

from mdance.cluster.nani import KmeansNANI

Data

• Load the data from a file, must be array of shape (n_samples, n_features).

• In this example, we will generate synthetic data using make_blobs from Fig. 2 of the NANI paper.

n_clusters = 7
data, true_labels = make_blobs(n_samples=1000, centers=n_clusters, n_features=2, random_state=0)

First, let’s checkout how state-of-the-art k-means performs on the data. which uses k-means++ initialization.

og_kmeans = KMeans(n_clusters=n_clusters, init='k-means++', n_init=1, random_state=42)
og_kmeans.fit(data)
og_kmeans_labels = og_kmeans.labels_

Visualize the clustered results to true labels.

fig1, ax1 = plt.subplots(1, 2, figsize=(12, 8), sharex=True, sharey=True)
ax1[0].scatter(data[:, 0], data[:, 1], c=og_kmeans_labels, cmap='tab10', s=20)
ax1[0].set_title('k-means++ Labels', fontsize=16, fontweight='bold')
ax1[1].scatter(data[:, 0], data[:, 1], c=true_labels, cmap='tab10', s=20)
ax1[1].set_title('True Labels', fontsize=16, fontweight='bold')
plt.show()

NANI

As you can see, k-means++ initialization did not get it right. Let’s use NANI to optimize initial centroids.

• Create an instance of KmeansNANI.

• data: data to cluster.

• n_clusters: number of clusters.

mod = KmeansNANI(data=data, n_clusters=n_clusters, metric='MSD', N_atoms=1,
                 init_type='strat_all', percentage=10)
initiators = mod.initiate_kmeans()
initiators = initiators[:n_clusters]

k-means with NANI

• Create an instance of KMeans.

• n_clusters: number of clusters.

• init: initial centroids.

• n_init: NANI only needs one initialization!

• random_state: We don’t need this because NANI is 100% deterministic!

kmeans = KMeans(n_clusters=n_clusters, init=initiators, n_init=1, random_state=None)
kmeans.fit(data)
kmeans_labels = kmeans.labels_

Plot

• Visualize the clustered results to true labels.

fig, ax2 = plt.subplots(1, 2, figsize=(12, 8), sharex=True, sharey=True)
ax2[0].scatter(data[:, 0], data[:, 1], c=kmeans_labels, cmap='tab10', s=20)
ax2[0].set_title('NANI Labels', fontsize=16, fontweight='bold')
ax2[1].scatter(data[:, 0], data[:, 1], c=true_labels, cmap='tab10', s=20)
ax2[1].set_title('True Labels', fontsize=16, fontweight='bold')
plt.show()

As you can see, NANI clustered the data perfectly!

That’s it! You have successfully used NANI to optimize initial centroids for k-means clustering.

• kmeans_labels: cluster labels assigned by k-means using NANI.

For more advance usage, please look at the NANI Tutorial. Why? Because NANI can also predict number of clusters, work with Molecular Dynamics data, and more!