kmeans.py 文件源码

def compute_db_index(matrix, kmeans):
    '''
    Compute Davies-Bouldin index, a measure of clustering quality.
    Faster and possibly more reliable than silhouette score.
    '''
    (n, m) = matrix.shape
    k = kmeans.n_clusters

    centers = kmeans.cluster_centers_
    labels = kmeans.labels_

    centroid_dists = sp_dist.squareform(sp_dist.pdist(centers))
    # Avoid divide-by-zero
    centroid_dists[np.abs(centroid_dists) < MIN_CENTROID_DIST] = MIN_CENTROID_DIST

    wss = np.zeros(k)
    counts = np.zeros(k)

    for i in xrange(n):
        label = labels[i]
        # note: this is 2x faster than scipy sqeuclidean
        sqdist = np.square(matrix[i,:] - centers[label,:]).sum()
        wss[label] += sqdist
        counts[label] += 1

    # Handle empty clusters
    counts[counts == 0] = 1

    scatter = np.sqrt(wss / counts)
    mixitude = (scatter + scatter[:, np.newaxis]) / centroid_dists
    np.fill_diagonal(mixitude, 0.0)

    worst_case_mixitude = np.max(mixitude, axis=1)
    db_score = worst_case_mixitude.sum() / k

    return db_score