kmedoid.py

from numpy.random import choice

class KMedoid:
    def __init__(self, k, verbose=False):
        self.k = k
        self.maxiter = 100
        self.verbose = verbose
        
    def assign_labels(self, S):
        return np.argmin(S, axis=1)
    
    def init_medoids(self, N):
        return choice(N, size=self.k, replace=False)
    
    def update_medoids(self, X, medoids):
        S = X[:, medoids]
        labels = self.assign_labels(S)
        new_medoids = []
        for l in set(labels):
            members = np.argwhere(labels == l).flatten() 
            s = X[:, members][members, :]
            m = np.argmin(np.sum(s, axis=0))
            new_medoids.append(members[m])
        return new_medoids
    
    def fit(self, X):
        medoids = self.init_medoids(len(X))
        if self.verbose:
            print(f'initial medoids: {medoids}')
            print(f'initial mean distance to the medoids: {np.mean(X[:, medoids])}')
            print()
        
        for i in range(self.maxiter):
            new_medoids = self.update_medoids(X, medoids)
            if self.verbose:
                print(f'step {i} medoids: {new_medoids}')
                print(f'step {i} mean distance to the medoids: {np.mean(X[:, new_medoids])}')
                print()
            
            if set(new_medoids) == set(medoids):
                break
            medoids = new_medoids
            
        return medoids