cluster/kmeans.py

from dataclasses import dataclass
import numpy as np

from cluster.clusterer import Clusterer


@dataclass
class Kmeans(Clusterer):
    k: int
    max_iter: int
    centroids = None
    clusters = None

    def build(
        self,
        X: np.array,
    ) -> None:
        # randomly initialize centroids
        centroids = X[np.random.choice(
            X.shape[0],
            self.k,
            replace=False,
        )]

        # Calculate Euclidean distance between each data point and each centroid
        # then assign each point to its closest cluster
        clusters = self.assign_clusters(X, centroids)
        centroids = self.update_centroids(self.k, X, clusters)

        while True:
            new_clusts = self.assign_clusters(X, centroids)
            if np.array_equal(new_clusts, clusters):
                break
            clusters = new_clusts
            centroids = self.update_centroids(self.k, X, clusters)

        self.clusters = clusters
        self.centroids = centroids

    @staticmethod
    def assign_clusters(
        X: np.array,
        centroids: np.array,
    ) -> np.array:
        distances = np.sqrt(((X - centroids[:, np.newaxis])**2).sum(axis=2))
        clusts = np.argmin(distances, axis=0)
        return clusts

    @staticmethod
    def update_centroids(
        k: int,
        X: np.array,
        clusters: np.array,
    ) -> np.array:
        centroids = np.zeros((k, X.shape[1]))
        for i in range(k):
            centroids[i] = X[clusters == i].mean(axis=0)
        return centroids

    def to_dict(
        self,
        X: np.array,
    ) -> dict:
        cluster_data = []
        for i in range(self.k):
            indices = np.where(self.clusters == i)[0]
            cluster_pts = X[indices].tolist()
            cluster_data.append({
                "cluster_id": i,
                "centroid": self.centroids[i].tolist(),
                "points": cluster_pts,
            })
        return {
            "k": self.k,
            "max_iter": self.max_iter,
            "clusters": cluster_data,
        }