I am implementing my own kmeans algorithm on a set of data. When I go with selecting any random points in the dataset as centroids, I am getting a very poor accuracy. But, when I go with selecting one centroid randomly from each class of data, I get a good accuracy. Please help me with where I am going wrong. Below is my implementation:

First, I generate random centroids and feed it to a function, to assign each point to a cluster based on which centroid it is closer to

def assignClustersKNN(features,centroids,labels):
    assignments = defaultdict(list)
    distances = [[0 for x in range(len(centroids))] for y in range(len(features))]
    #Iterating over all data points
    for i in range(len(features)):
        #Iterating over all centroids
        for j in range(0,len(centroids)):
            distances[i][j] = euclidean(features[i],centroids[j])
        #Getting the index of the centroid which is the closest
        clusterAssigned = distances[i].index(min(distances[i]))
        #adding the point to the closest cluster
        assignments[clusterAssigned].append(features[i])    
    return assignments 

Then, I update the centroid of each cluster by computing the mean of the points in a cluster, which is the centroid of that cluster

def updateCentroids(assignments):
    newCentroids = np.zeros(shape=(len(assignments.keys()),3))
    for i in assignments.keys():
        #getting the datapoints of each cluster
        clusterMembers = assignments[i]
        #computing the mean of the datapoints of the cluster
        newCentroids[i] = np.mean(clusterMembers,axis=0)
    return newCentroids    

I have chosen the stopping condition as , when the centroids of a cluster in an iteration do not differ from the centroids from the previous iteration, that means that the clusters have not changed and I stop the process

0 Answers