Free Access
Table 3.
Hyperparameters we explored for each algorithm.
Hyperparameters | Tested values | Best value |
---|---|---|
Affinity propagation | ||
affinity | euclidean | euclidean |
convergence_iter | 20 | 20 |
damping | [0.5, 0.51,..., 0.99] | 0.5 |
max_iter | 200 | 200 |
Agglomerative clustering | ||
affinity | [manhattan, cosine, | |
euclidean] | manhattan | |
compute_full_tree | [False, True] | False |
linkage | [complete, average] | complete |
n_clusters | [9, 10,..., 26] | 26 |
Bayesian Gaussian mixture | ||
covariance_type | [tied, diag] | diag |
init_params | [kmeans, random] | kmeans |
n_components | [9, 10,..., 26] | 20 |
n_init | 5 | 5 |
random_state ( † ) | [31, 43,..., 473] | 283 |
warm_start | [False, True] | False |
DBSCAN | ||
algorithm | [ball_tree, kd_tree, | |
brute] | ball_tree | |
eps | [0.3, 0.31,..., 1.49] | 0.58 |
leaf_size | 10 | 10 |
metric | euclidean | euclidean |
min_samples | [2, 3] | 2 |
Gaussian mixture | ||
covariance_type | [spherical, diag] | spherical |
init_params | [k-means, random] | k-means |
max_iter | 1000 | 1000 |
n_components | [2, 3,..., 26] | 20 |
random_state† | [31, 43,..., 473] | 283 |
K-means | ||
init | [k-means++, random] | random |
n_clusters | [9, 10,..., 26] | 26 |
n_init | 5 | 5 |
random_state ( † ) | [31, 43,..., 473] | 43 |
Mini-batch K-means | ||
batch_size | [10, 20,..., 100] | 90 |
init | [k-means++, random] | random |
n_clusters | [9, 10,..., 26] | 25 |
random_state ( † ) | [31, 43,..., 473] | 31 |
reassignment_ratio | [0.1, 0.01, 0.001] | 0.1 |
Spectral clustering | ||
affinity | [rbf, sigmoid, | |
polynomial, poly] | rbf | |
assign_labels | [kmeans, discretize] | discretize |
degree | [3, 4, 5] | 3 |
n_clusters | [9, 10,..., 26] | 17 |
n_neighbors | [2, 5, 10] | 2 |
random_state ( † ) | [31, 43,..., 473] | 383 |
Notes. The last column shows the value of the parameter in the run with the highest homogeneity score.
( † )We have tried ten different random seeds, 31, 43, 98, 196, 283, 294, 374, 383, 433 and 473.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.