def test_experiment_mcc():
mat_obj = sio.loadmat("../data/mcc/chess.mat")
x = mat_obj['input_space']
y = mat_obj['target_space'] ## this is not one hot for scc
y = y.reshape(-1)
wrange = []
for u in range(x.shape[0]):
if 2**u <= x.shape[0]:
wrange.append(int(2**u))
wrange.append(int(x.shape[0]))
relief_b3_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True,
use_average_neighbour=True,
determine_k_automatically=True,
num_iter=wrange)
relief_b3_instance.fit(x, y)
relief_b3_instance.feature_importances_
all_weights = relief_b3_instance.all_weights
print(all_weights)
for k, v in relief_b3_instance.timed.items():
print(k + "\t" + str(v))
def test_experiment_arff_mlc():
ova = "../data/test.arff"
features = [0, 1, 2, 3, 4]
targets = [5, 6]
x_train, y_train, _ = load_arff(ova, features, targets)
neighbours = 2
print("ReliefF ..")
relief_b3_instance = reliefe.ReliefE(embedding_based_distances=False,
verbose=True,
k=neighbours)
relief_b3_instance.fit(x_train, y_train)
importances2 = relief_b3_instance.feature_importances_
print(importances2)
print("ReliefE ..")
relief_b3_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True,
k=neighbours,
latent_dimension=2)
relief_b3_instance.fit(x_train, y_train)
importances2 = relief_b3_instance.feature_importances_
print(importances2)
x_train, y_train, _ = reliefe.utils.load_arff(ova, features, targets[0])
neighbours = 2
print("ReliefF ..")
relief_b3_instance = reliefe.ReliefE(embedding_based_distances=False,
verbose=True,
k=neighbours)
relief_b3_instance.fit(x_train, y_train)
importances2 = relief_b3_instance.feature_importances_
print(importances2)
print("ReliefE ..")
relief_b3_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True,
k=neighbours,
latent_dimension=2)
relief_b3_instance.fit(x_train, y_train)
importances2 = relief_b3_instance.feature_importances_
print(importances2)
def test_simple_benchmark():
sns.set_style("whitegrid")
data_obj = load_breast_cancer()
x = data_obj['data']
y = data_obj['target']
names = data_obj['feature_names']
# let's overfit, just for demo purposes
reliefE_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True,
use_average_neighbour=False,
determine_k_automatically=False,
mlc_distance="hamming")
reliefE_instance.fit(x, y)
reliefe_importances = reliefE_instance.feature_importances_
rf_model = RandomForestClassifier()
rf_model.fit(x, y)
rf = rf_model.feature_importances_
mutual_information = mutual_info_classif(x, y)
sorted_rf = np.argsort(rf)
sorted_mi = np.argsort(mutual_information)
sorted_re = np.argsort(reliefe_importances)
names = ["RF", "MI", "ReliefE"]
indices = [sorted_rf, sorted_mi, sorted_re]
output_struct = {}
for name, indice_set in zip(names, indices):
scores = []
indice_set = indice_set.tolist()
print("Computing evaluations for: {}".format(name))
for j in tqdm.tqdm(range(len(indice_set))):
selected_features = indice_set[0:j + 1]
subset = x[:, selected_features]
clf = LogisticRegression(max_iter=10000000, solver="lbfgs")
score = np.mean(cross_val_score(clf, subset, y, cv=10))
scores.append((score, j + 1))
output_struct[name] = scores
print("Plotting ..")
for k, v in output_struct.items():
indices = []
scores = []
for x, y in v:
indices.append(y)
scores.append(x)
plt.plot(indices, scores, label=k)
plt.xlabel("Top features")
plt.ylabel("Performance (Accuracy)")
plt.legend()
plt.show()
def test_experiment_mlc_autok(dataset):
mat_obj = sio.loadmat(dataset)
x = mat_obj['input_space']
y = mat_obj['target_space'] ## this is not one hot for scc
assert y.shape[1] > 1
reliefe_instance = reliefe.ReliefE(embedding_based_distances=False,
verbose=True,
use_average_neighbour=False,
determine_k_automatically=True,
num_iter=50)
reliefe_instance.fit(x, y)
assert len(reliefe_instance.feature_importances_) > 0
def test_custom_embedding():
mat_obj = sio.loadmat("../data/mcc/chess.mat")
x = mat_obj['input_space']
y = mat_obj['target_space'] ## this is not one hot for scc
y = y.reshape(-1)
reliefe_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True,
use_average_neighbour=False,
determine_k_automatically=False,
num_iter=50)
emb_custom = PCA()
reliefe_instance.fit(x, y, embedding_method=emb_custom)
assert len(reliefe_instance.feature_importances_) > 0
import reliefe
import scipy.io as sio
# Load the data first
mat_obj = sio.loadmat("../data/mlc/Science1.mat")
x = mat_obj['input_space'] ## scipy csr sparse matrix (or numpy dense)
y = mat_obj['target_space'] ## scipy csr sparse matrix (or numpy dense)
reliefE_instance = reliefe.ReliefE(determine_k_automatically=True,
embedding_based_distances=True,
num_iter=128,
verbose=True)
# Initialize default ReliefE
reliefE_instance.fit(x, y) # Compute rankings
print(reliefE_instance.feature_importances_
) # rankings for features (same order as x)
reliefE_instance = reliefe.ReliefE(determine_k_automatically=True,
num_iter=128,
verbose=True)
# Initialize default ReliefE
reliefE_instance.fit(x, y) # Compute rankings
print(reliefE_instance.feature_importances_
) # rankings for features (same order as x)
import reliefe
import scipy.io as sio
from sklearn.decomposition import TruncatedSVD
mat_obj = sio.loadmat("../data/mcc/chess.mat")
x = mat_obj['input_space']
y = mat_obj['target_space'] ## this is not one hot for scc
y = y.reshape(-1)
print(x.shape, y.shape)
reliefe_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True)
# Simply provide a sklearn-like transform object
emb_custom = TruncatedSVD() # Let's do SVD
# Provide it as the "embedding_method" parameter
reliefe_instance.fit(x, y, embedding_method=emb_custom)
assert len(reliefe_instance.feature_importances_) > 0
print(reliefe_instance.feature_importances_)
def test_importances(importances, method="ReliefE"):
sorted_importances = np.argsort(importances)[::-1]
## If the first two features are amongst the top 3, this is a relatively OK sign.
if 0 in sorted_importances[0:3] and 1 in sorted_importances[0:3]:
print("Successfully retrieved top features.")
return 1, method
else:
return 0, method
df = pd.read_csv("synthetic.txt", sep=",")
print(df)
## Features a1 and a2 are here crucial for performance.
x = df[[f"a{x}" for x in range(1, 9)]].values
y = df.Class.astype(int).values
## ReliefE will revert to the core version in LD settings; the point of reducing the dimension makes no sense in LD; the core version is fast for this type of problems.
reliefE_instance = reliefe.ReliefE(
embedding_based_distances=True,
num_iter=0.05, ## Few-shot?
verbose=True,
use_average_neighbour=True,
determine_k_automatically=True)
reliefE_instance.fit(x, y)
importances = reliefE_instance.feature_importances_
test_importances(importances, method=f"ReliefE")
import scipy.io as sio
import reliefe
mat_obj = sio.loadmat("../data/mlc/medical.mat")
x = mat_obj['input_space']
y = mat_obj['target_space'] ## this is not one hot for scc
wrange = []
# Fully fledged MLC - ReliefE (with all functionality)
reliefE_instance = reliefe.ReliefE(embedding_based_distances=True,
verbose=True,
use_average_neighbour=False,
determine_k_automatically=False,
mlc_distance="hamming")
reliefE_instance.fit(x, y)
print(reliefE_instance.feature_importances_)