models.py

import numpy as np

class KMeans:
  # Initialization and properties
  def __init__(self):
    self.centroids = np.empty(1)

  # Fit method
  def fit(self, data, clusters, epochs = 1, random_seed = 42):
    np.random.seed(random_seed)
    N = len(data)
    centroids = data[np.random.choice(N, clusters, replace=False), :]

    labels = np.empty(N)
    old_labels = np.empty(N)

    while True:
      distances = np.linalg.norm(data[:, None, :] - centroids, axis=2)
      labels = np.argmin(distances, axis=1)
      for j in range(clusters):
            centroids[j] = np.mean(data[labels == j], axis=0)

      if np.all(labels == old_labels):
            break

      old_labels = labels
    self.centroids = centroids
    return labels

  # Predict method
  def predict(self, data):
      distances = np.linalg.norm(data[:, None, :] - self.centroids, axis=2)
      labels = np.argmin(distances, axis=1)
      return labels
  

class KMedoids:
  # Initialization and properties
  def __init__(self, data, clusters):
    self.medoids = np.empty(1)
    self.data = data
    self.N = len(data)
    self.clusters = clusters

  # Fit
  def fit(self, random_seed = 42):
    np.random.seed(random_seed)
    data = self.data
    N = self.N
    clusters = self.clusters
    medoids_idx = np.random.choice(N, clusters, replace=False)
    medoids = data[medoids_idx].copy()
    distances = np.zeros((N, clusters))

    for i in range(clusters):
        distances[:, i] = np.sum(np.abs(data - medoids[i]), axis=1)

    labels = np.argmin(distances, axis=1)
    old_labels = np.empty(N)
    all_idxs = np.arange(N)

    while True:
      best_swap = (-1, -1, 0)
      best_distances = np.zeros(N)
      for i in range(clusters):
          non_medoids_idx = all_idxs[np.logical_not(np.isin(all_idxs, medoids_idx))]
          for j in non_medoids_idx:
              new_medoid = data[j]
              new_distances = np.sum(np.abs(data - new_medoid), axis=1)
              cost_change = np.sum(new_distances[labels == i]) - np.sum(
                  distances[labels == i, i]
              )
              if cost_change < best_swap[2]:
                  best_swap = (i, j, cost_change)
                  best_distances = new_distances

      if best_swap == (-1, -1, 0):
            break
      i, j, _ = best_swap
      distances[:, i] = best_distances
      medoids[i] = data[j]

      labels = np.argmin(distances, axis=1)

      old_labels = labels
    self.medoids = medoids
    return labels

  # Predict
  def predict(self, data):
      distances = np.zeros((len(data), self.clusters))
      for i in range(self.clusters):
        distances[:, i] = np.sum(np.abs(data - self.medoids[i]), axis=1)
      labels = np.argmin(distances, axis=1)
      return labels
  
  
class EnsembleClustering:
  # Initialization
  def __init__(self, data, clusters):
    self.data = data
    self.clusters = clusters
    self.kmeans = None
    self.kmedoids = None

  # Fit method
  def fit(self):
    kmeans = KMeans()
    kmeans_labels = kmeans.fit(self.data, self.clusters)
    self.kmeans = kmeans

    kmedoids = KMedoids(data = self.data, clusters = self.clusters)
    kmedoids_labels = kmedoids.fit()
    self.kmedoids = kmedoids

    labels = self.maximumVoting(kmeans_labels, kmedoids_labels)
    return labels

  # Maximum voting method
  def maximumVoting(self, labels1, labels2):
    labels = np.zeros(len(labels1), dtype=int)
    for i in range(len(labels1)):
      voting = np.zeros(self.clusters, dtype=int)
      voting[labels1[i]] += 1
      voting[labels2[i]] += 1
      labels[i] = voting.argmax()
    return labels

  # Predict method
  def predict(self, data):
    kmeans_labels = self.kmeans.predict(data)
    kmedoids_labels = self.kmedoids.predict(data)
    labels = self.maximumVoting(kmeans_labels, kmedoids_labels)
    return labels