def main():
# Read the result from the previous chapter convert the index to datetime
try:
dataset = pd.read_csv(DATA_PATH / DATASET_FNAME, index_col=0)
dataset.index = pd.to_datetime(dataset.index)
except IOError as e:
print(
'File not found, try to run previous crowdsignals scripts first!')
raise e
# Create an instance of visualization class to plot the results
DataViz = VisualizeDataset(__file__)
# Create objects for clustering
clusteringNH = NonHierarchicalClustering()
clusteringH = HierarchicalClustering()
if FLAGS.mode == 'kmeans':
# Do some initial runs to determine the right number for k
k_values = range(2, 10)
silhouette_values = []
print('Running k-means clustering')
for k in k_values:
print(f'k = {k}')
dataset_cluster = clusteringNH.k_means_over_instances(
dataset=copy.deepcopy(dataset),
cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
k=k,
distance_metric='default',
max_iters=20,
n_inits=10)
silhouette_score = dataset_cluster['silhouette'].mean()
print(f'silhouette = {silhouette_score}')
silhouette_values.append(silhouette_score)
DataViz.plot_xy(x=[k_values],
y=[silhouette_values],
xlabel='k',
ylabel='silhouette score',
ylim=[0, 1],
line_styles=['b-'])
# Run the knn with the highest silhouette score
k = k_values[np.argmax(silhouette_values)]
print(f'Highest K-Means silhouette score: k = {k}')
print('Use this value of k to run the --mode=final --k=?')
if FLAGS.mode == 'kmediods':
# Do some initial runs to determine the right number for k
k_values = range(2, 10)
silhouette_values = []
print('Running k-medoids clustering')
for k in k_values:
print(f'k = {k}')
dataset_cluster = clusteringNH.k_medoids_over_instances(
dataset=copy.deepcopy(dataset),
cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
k=k,
distance_metric='default',
max_iters=20,
n_inits=10)
silhouette_score = dataset_cluster['silhouette'].mean()
print(f'silhouette = {silhouette_score}')
silhouette_values.append(silhouette_score)
DataViz.plot_xy(x=[k_values],
y=[silhouette_values],
xlabel='k',
ylabel='silhouette score',
ylim=[0, 1],
line_styles=['b-'])
# Run k medoids with the highest silhouette score
k = k_values[np.argmax(silhouette_values)]
print(f'Highest K-Medoids silhouette score: k = {k}')
dataset_kmed = clusteringNH.k_medoids_over_instances(
dataset=copy.deepcopy(dataset),
cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
k=k,
distance_metric='default',
max_iters=20,
n_inits=50)
DataViz.plot_clusters_3d(
data_table=dataset_kmed,
data_cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
cluster_col='cluster',
label_cols=['label'])
DataViz.plot_silhouette(data_table=dataset_kmed,
cluster_col='cluster',
silhouette_col='silhouette')
util.print_latex_statistics_clusters(
dataset=dataset_kmed,
cluster_col='cluster',
input_cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
label_col='label')
# Run hierarchical clustering
if FLAGS.mode == 'agglomerative':
k_values = range(2, 10)
silhouette_values = []
# Do some initial runs to determine the right number for the maximum number of clusters
print('Running agglomerative clustering')
for k in k_values:
print(f'k = {k}')
dataset_cluster, link = clusteringH.agglomerative_over_instances(
dataset=dataset,
cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
max_clusters=k,
distance_metric='euclidean',
use_prev_linkage=True,
link_function='ward')
silhouette_score = dataset_cluster['silhouette'].mean()
print(f'silhouette = {silhouette_score}')
silhouette_values.append(silhouette_score)
if k == k_values[0]:
DataViz.plot_dendrogram(dataset_cluster, link)
# Plot the clustering results
DataViz.plot_xy(x=[k_values],
y=[silhouette_values],
xlabel='k',
ylabel='silhouette score',
ylim=[0, 1],
line_styles=['b-'])
if FLAGS.mode == 'final':
# Select the outcome dataset of the knn clustering
clusteringNH = NonHierarchicalClustering()
dataset = clusteringNH.k_means_over_instances(
dataset=dataset,
cols=['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
k=FLAGS.k,
distance_metric='default',
max_iters=50,
n_inits=50)
# Plot the results
DataViz.plot_clusters_3d(dataset,
['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
'cluster', ['label'])
DataViz.plot_silhouette(dataset, 'cluster', 'silhouette')
# Print table statistics
util.print_latex_statistics_clusters(
dataset, 'cluster', ['acc_phone_x', 'acc_phone_y', 'acc_phone_z'],
'label')
del dataset['silhouette']
# Store the final dataset
dataset.to_csv(DATA_PATH / RESULT_FNAME)
print("date and time =", dt_string)
diff = now3 - now2
print('difference time', diff)
# First, let us consider the performance over a selection of features:
fs = FeatureSelectionClassification()
features, ordered_features, ordered_scores = fs.forward_selection(
N_FORWARD_SELECTION, train_X[features_after_chapter_5], train_y)
print(ordered_scores)
print(ordered_features)
DataViz.plot_xy(x=[range(1, N_FORWARD_SELECTION + 1)],
y=[ordered_scores],
xlabel='number of features',
ylabel='accuracy')
# datetime object containing current date and time
now4 = datetime.now()
# dd/mm/YY H:M:S
dt_string = now4.strftime("%d/%m/%Y %H:%M:%S")
print("date and time =", dt_string)
diff = now4 - now3
print('difference time', diff)
# Based on the plot we select the top 10 features (note: slightly different compared to Python 2, we use
# those feartures here).
selected_features = [
dataset = LowPass.low_pass_filter(dataset, col, fs, cutoff, order=10)
DataViz.plot_dataset(dataset.iloc[int(0.5 * len(new_dataset.index)):int(0.8 * len(new_dataset.index))],
[col, col + '_lowpass'], ['exact', 'exact'], ['line', 'line'])
dataset[col] = dataset[col + '_lowpass']
del dataset[col + '_lowpass']
# Determine the PC's for all but our target columns (the labels and the heart rate)
# We simplify by ignoring both, we could also ignore one first, and apply a PC to the remainder.
PCA = PrincipalComponentAnalysis()
selected_predictor_cols = [c for c in dataset.columns if (not ('label' in c))]
pc_values = PCA.determine_pc_explained_variance(dataset, selected_predictor_cols)
# Plot the variance explained.
DataViz.plot_xy(x=[range(1, len(selected_predictor_cols) + 1)], y=[pc_values],
xlabel='principal component number', ylabel='explained variance',
ylim=[0, 1], line_styles=['b-'])
# We select 7 as the best number of PC's as this explains most of the variance
n_pcs = 3
dataset = PCA.apply_pca(copy.deepcopy(dataset), selected_predictor_cols, n_pcs)
# And we visualize the result of the PC's
DataViz.plot_dataset(dataset, ['pca_', 'label'], ['like', 'like'], ['line', 'points'])
# And the overall final dataset:
DataViz.plot_dataset(dataset,
'Gyroscope z (rad/s)'
]
## Do some initial runs to determine the right number for k
print('===== kmeans clustering =====')
for k in k_values:
print(f'k = {k}')
dataset_cluster = clusteringNH.k_means_over_instances(
copy.deepcopy(dataset), attributes_to_cluster, k, 'default', 20, 10)
silhouette_score = dataset_cluster['silhouette'].mean()
print(f'silhouette = {silhouette_score}')
silhouette_values.append(silhouette_score)
DataViz.plot_xy(x=[k_values],
y=[silhouette_values],
xlabel='k',
ylabel='silhouette score',
ylim=[0, 1],
line_styles=['b-'])
# And run the knn with the highest silhouette score
# k = 6 # todo: replaced with np.argmax call over silhouette scores
k = k_values[np.argmax(silhouette_values)]
print(f'Highest K-Means silhouette score: k = {k}')
dataset_knn = clusteringNH.k_means_over_instances(copy.deepcopy(dataset),
attributes_to_cluster, k,
'default', 50, 50)
DataViz.plot_clusters_3d(dataset_knn, attributes_to_cluster, 'cluster',
['label'])
DataViz.plot_silhouette(dataset_knn, 'cluster', 'silhouette')
def main():
# Read the result from the previous chapter and convert the index to datetime
try:
dataset = pd.read_csv(DATA_PATH / DATASET_FILENAME, index_col=0)
dataset.index = pd.to_datetime(dataset.index)
except IOError as e:
print(
'File not found, try to run previous crowdsignals scripts first!')
raise e
# Create an instance of visualization class to plot the results
DataViz = VisualizeDataset(__file__)
# Consider the first task, namely the prediction of the label. Therefore create a single column with the categorical
# attribute representing the class. Furthermore, use 70% of the data for training and the remaining 30% as an
# independent test set. Select the sets based on stratified sampling and remove cases where the label is unknown.
print('\n- - - Loading dataset - - -')
prepare = PrepareDatasetForLearning()
learner = ClassificationAlgorithms()
evaluation = ClassificationEvaluation()
train_X, test_X, train_y, test_y = prepare.split_single_dataset_classification(
dataset, ['label'], 'like', 0.7, filter_data=True, temporal=False)
print('Training set length is: ', len(train_X.index))
print('Test set length is: ', len(test_X.index))
# Select subsets of the features
print('- - - Selecting subsets - - -')
basic_features = [
'acc_phone_x', 'acc_phone_y', 'acc_phone_z', 'acc_watch_x',
'acc_watch_y', 'acc_watch_z', 'gyr_phone_x', 'gyr_phone_y',
'gyr_phone_z', 'gyr_watch_x', 'gyr_watch_y', 'gyr_watch_z',
'hr_watch_rate', 'light_phone_lux', 'mag_phone_x', 'mag_phone_y',
'mag_phone_z', 'mag_watch_x', 'mag_watch_y', 'mag_watch_z',
'press_phone_pressure'
]
pca_features = [
'pca_1', 'pca_2', 'pca_3', 'pca_4', 'pca_5', 'pca_6', 'pca_7'
]
time_features = [name for name in dataset.columns if '_temp_' in name]
freq_features = [
name for name in dataset.columns
if (('_freq' in name) or ('_pse' in name))
]
cluster_features = ['cluster']
print('#basic features: ', len(basic_features))
print('#PCA features: ', len(pca_features))
print('#time features: ', len(time_features))
print('#frequency features: ', len(freq_features))
print('#cluster features: ', len(cluster_features))
features_after_chapter_3 = list(set().union(basic_features, pca_features))
features_after_chapter_4 = list(set().union(features_after_chapter_3,
time_features, freq_features))
features_after_chapter_5 = list(set().union(features_after_chapter_4,
cluster_features))
if FLAGS.mode == 'selection' or FLAGS.mode == 'all':
# First, consider the performance over a selection of features
N_FORWARD_SELECTION = FLAGS.nfeatures
fs = FeatureSelectionClassification()
print('\n- - - Running feature selection - - -')
features, ordered_features, ordered_scores = fs.forward_selection(
max_features=N_FORWARD_SELECTION,
X_train=train_X[features_after_chapter_5],
y_train=train_y)
DataViz.plot_xy(x=[range(1, N_FORWARD_SELECTION + 1)],
y=[ordered_scores],
xlabel='number of features',
ylabel='accuracy')
# Select the most important features (based on python2 features)
selected_features = [
'acc_phone_y_freq_0.0_Hz_ws_40',
'press_phone_pressure_temp_mean_ws_120', 'gyr_phone_x_temp_std_ws_120',
'mag_watch_y_pse', 'mag_phone_z_max_freq', 'gyr_watch_y_freq_weighted',
'gyr_phone_y_freq_1.0_Hz_ws_40', 'acc_phone_x_freq_1.9_Hz_ws_40',
'mag_watch_z_freq_0.9_Hz_ws_40', 'acc_watch_y_freq_0.5_Hz_ws_40'
]
if FLAGS.mode == 'regularization' or FLAGS.mode == 'all':
print('\n- - - Running regularization and model complexity test - - -')
# Study the impact of regularization and model complexity: does regularization prevent overfitting?
# Due to runtime constraints run the experiment 3 times, for even more robust data increase the repetitions
N_REPEATS_NN = FLAGS.nnrepeat
reg_parameters = [0.0001, 0.001, 0.01, 0.1, 1, 10]
performance_training = []
performance_test = []
for reg_param in reg_parameters:
performance_tr = 0
performance_te = 0
for i in range(0, N_REPEATS_NN):
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.feedforward_neural_network(
train_X,
train_y,
test_X,
hidden_layer_sizes=(250, ),
alpha=reg_param,
max_iter=500,
gridsearch=False)
performance_tr += evaluation.accuracy(train_y, class_train_y)
performance_te += evaluation.accuracy(test_y, class_test_y)
performance_training.append(performance_tr / N_REPEATS_NN)
performance_test.append(performance_te / N_REPEATS_NN)
DataViz.plot_xy(x=[reg_parameters, reg_parameters],
y=[performance_training, performance_test],
method='semilogx',
xlabel='regularization parameter value',
ylabel='accuracy',
ylim=[0.95, 1.01],
names=['training', 'test'],
line_styles=['r-', 'b:'])
if FLAGS.mode == 'tree' or FLAGS.mode == 'all':
print('\n- - - Running leaf size test of decision tree - - -')
# Consider the influence of certain parameter settings for the tree model. (very related to the
# regularization) and study the impact on performance.
leaf_settings = [1, 2, 5, 10]
performance_training = []
performance_test = []
for no_points_leaf in leaf_settings:
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.decision_tree(
train_X[selected_features],
train_y,
test_X[selected_features],
min_samples_leaf=no_points_leaf,
gridsearch=False,
print_model_details=False)
performance_training.append(
evaluation.accuracy(train_y, class_train_y))
performance_test.append(evaluation.accuracy(test_y, class_test_y))
DataViz.plot_xy(x=[leaf_settings, leaf_settings],
y=[performance_training, performance_test],
xlabel='Minimum number of points per leaf',
ylabel='Accuracy',
names=['training', 'test'],
line_styles=['r-', 'b:'])
if FLAGS.mode == 'overall' or FLAGS.mode == 'all':
print(
'\n- - - Running test of all different classification algorithms - - -'
)
# Perform grid searches over the most important parameters and do so by means of cross validation upon the
# training set
possible_feature_sets = [
basic_features, features_after_chapter_3, features_after_chapter_4,
features_after_chapter_5, selected_features
]
feature_names = [
'initial set', 'Chapter 3', 'Chapter 4', 'Chapter 5',
'Selected features'
]
N_KCV_REPEATS = FLAGS.kcvrepeat
scores_over_all_algs = []
for i in range(0, len(possible_feature_sets)):
selected_train_X = train_X[possible_feature_sets[i]]
selected_test_X = test_X[possible_feature_sets[i]]
# First run non deterministic classifiers a number of times to average their score
performance_tr_nn, performance_te_nn = 0, 0
performance_tr_rf, performance_te_rf = 0, 0
performance_tr_svm, performance_te_svm = 0, 0
for repeat in range(0, N_KCV_REPEATS):
print(
f'Training NeuralNetwork run {repeat + 1} / {N_KCV_REPEATS}, featureset is {feature_names[i]} ... '
)
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.feedforward_neural_network(
selected_train_X,
train_y,
selected_test_X,
gridsearch=True)
print(
f'Training RandomForest run {repeat + 1} / {N_KCV_REPEATS}, featureset is {feature_names[i]} ... '
)
performance_tr_nn += evaluation.accuracy(
train_y, class_train_y)
performance_te_nn += evaluation.accuracy(test_y, class_test_y)
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.random_forest(
selected_train_X,
train_y,
selected_test_X,
gridsearch=True)
performance_tr_rf += evaluation.accuracy(
train_y, class_train_y)
performance_te_rf += evaluation.accuracy(test_y, class_test_y)
print(
f'Training SVM run {repeat + 1} / {N_KCV_REPEATS}, featureset is {feature_names[i]} ...'
)
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner. \
support_vector_machine_with_kernel(selected_train_X, train_y, selected_test_X, gridsearch=True)
performance_tr_svm += evaluation.accuracy(
train_y, class_train_y)
performance_te_svm += evaluation.accuracy(test_y, class_test_y)
overall_performance_tr_nn = performance_tr_nn / N_KCV_REPEATS
overall_performance_te_nn = performance_te_nn / N_KCV_REPEATS
overall_performance_tr_rf = performance_tr_rf / N_KCV_REPEATS
overall_performance_te_rf = performance_te_rf / N_KCV_REPEATS
overall_performance_tr_svm = performance_tr_svm / N_KCV_REPEATS
overall_performance_te_svm = performance_te_svm / N_KCV_REPEATS
# Run deterministic classifiers:
print("Deterministic Classifiers:")
print(
f'Training Nearest Neighbor run 1 / 1, featureset {feature_names[i]}'
)
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.k_nearest_neighbor(
selected_train_X, train_y, selected_test_X, gridsearch=True)
performance_tr_knn = evaluation.accuracy(train_y, class_train_y)
performance_te_knn = evaluation.accuracy(test_y, class_test_y)
print(
f'Training Decision Tree run 1 / 1 featureset {feature_names[i]}'
)
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.decision_tree(
selected_train_X, train_y, selected_test_X, gridsearch=True)
performance_tr_dt = evaluation.accuracy(train_y, class_train_y)
performance_te_dt = evaluation.accuracy(test_y, class_test_y)
print(
f'Training Naive Bayes run 1/1 featureset {feature_names[i]}')
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.naive_bayes(
selected_train_X, train_y, selected_test_X)
performance_tr_nb = evaluation.accuracy(train_y, class_train_y)
performance_te_nb = evaluation.accuracy(test_y, class_test_y)
scores_with_sd = util. \
print_table_row_performances(feature_names[i], len(selected_train_X.index),
len(selected_test_X.index), [
(overall_performance_tr_nn, overall_performance_te_nn),
(overall_performance_tr_rf, overall_performance_te_rf),
(overall_performance_tr_svm, overall_performance_te_svm),
(performance_tr_knn, performance_te_knn),
(performance_tr_knn, performance_te_knn),
(performance_tr_dt, performance_te_dt),
(performance_tr_nb, performance_te_nb)])
scores_over_all_algs.append(scores_with_sd)
DataViz.plot_performances_classification(
['NN', 'RF', 'SVM', 'KNN', 'DT', 'NB'], feature_names,
scores_over_all_algs)
if FLAGS.mode == 'detail' or FLAGS.mode == 'all':
print(
'\n- - - Running detail test of promising classification algorithms - - -'
)
# Study two promising ones in more detail, namely decision tree and random forest algorithm
learner.decision_tree(train_X[selected_features],
train_y,
test_X[selected_features],
gridsearch=True,
print_model_details=True,
export_tree_path=EXPORT_TREE_PATH)
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.random_forest(
train_X[selected_features],
train_y,
test_X[selected_features],
gridsearch=True,
print_model_details=True)
test_cm = evaluation.confusion_matrix(test_y, class_test_y,
class_train_prob_y.columns)
DataViz.plot_confusion_matrix(test_cm,
class_train_prob_y.columns,
normalize=False)
train_y,
test_X,
hidden_layer_sizes=(250, ),
alpha=reg_param,
max_iter=500,
gridsearch=False)
performance_tr += eval.accuracy(train_y, class_train_y)
performance_te += eval.accuracy(test_y, class_test_y)
performance_training.append(performance_tr / N_REPEATS_NN)
performance_test.append(performance_te / N_REPEATS_NN)
DataViz.plot_xy(x=[reg_parameters, reg_parameters],
y=[performance_training, performance_test],
method='semilogx',
xlabel='regularization parameter value',
ylabel='accuracy',
ylim=[0.95, 1.01],
names=['training', 'test'],
line_styles=['r-', 'b:'])
# Second, let us consider the influence of certain parameter settings for the tree model. (very related to the
# regularization) and study the impact on performance.
leaf_settings = [1, 2, 5, 10]
performance_training = []
performance_test = []
for no_points_leaf in leaf_settings:
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.decision_tree(
train_X[selected_features],
def experiment(file):
dataset = pd.read_csv(file, index_col=time_col)
DataViz = VisualizeDataset(__file__.split('.')[0] +
file.split('.')[0].split('/')[1] + '.py',
show=True)
print(DataViz.figures_dir)
dataset.index = pd.to_datetime(dataset.index)
prepare = PrepareDatasetForLearning()
train_X, test_X, train_y, test_y = prepare.split_single_dataset_classification(
dataset, ['label'],
'like',
0.7,
filter=False,
temporal=False,
drop_na=False,
fill_na=True)
time_features = [name for name in dataset.columns if '_temp' in name]
freq_features = [
name for name in dataset.columns
if (('_freq' in name) or ('_pse' in name))
]
cluster_features = ['cluster']
features_2 = list(set().union(basic_features, time_features))
features_3 = list(set().union(basic_features, time_features,
freq_features))
features_4 = list(set().union(basic_features, time_features, freq_features,
cluster_features))
# print('feature selection')
# fs = FeatureSelectionClassification()
# features, selected_features, ordered_scores = fs.forward_selection(N_FORWARD_SELECTION,
# train_X[features_4], train_y)
# log([str(ordered_scores), str(selected_features)])
selected_features = [
'gyr_y_temp_std_ws_1200', 'acc_z_temp_mean_ws_120',
'acc_x_temp_mean_ws_120', 'gyr_x_temp_std_ws_2400', 'gyr_z_max_freq',
'gyr_y_freq_1.9_Hz_ws_40', 'acc_z_freq_0.4_Hz_ws_40',
'gyr_z_freq_1.2_Hz_ws_40', 'gyr_x_freq_0.2_Hz_ws_40',
'acc_z_freq_1.0_Hz_ws_40', 'acc_x_freq_0.2_Hz_ws_40',
'acc_y_freq_1.9_Hz_ws_40', 'gyr_x_temp_mean_ws_1200',
'acc_z_freq_1.9_Hz_ws_40', 'acc_x_temp_std_ws_120',
'gyr_z_temp_std_ws_120', 'gyr_y_freq_1.5_Hz_ws_40',
'gyr_z_temp_mean_ws_120', 'gyr_x_freq_0.0_Hz_ws_40',
'acc_z_freq_0.6_Hz_ws_40'
]
DataViz.plot_xy(x=[range(1, N_FORWARD_SELECTION + 1)],
y=[selected_features],
xlabel='number of features',
ylabel='accuracy')
print('feature selection finished for %s' % file)
learner = ClassificationAlgorithms()
eval = ClassificationEvaluation()
possible_feature_sets = [
basic_features, features_2, features_3, features_4, selected_features
]
feature_names = [
'Basic features', 'Features with time', 'Features with frequency',
'Features with cluster', 'Selected features'
]
# with shelve.open('temp/shelve.out', 'n') as f:
# for key in dir():
# try:
# f[key] = globals()[key]
# except:
# print('ERROR shelving: {0}'.format(key))
N_KCV_REPEATS = 1
scores_over_all_algs = []
for i in range(0, len(possible_feature_sets)):
print(datetime.now())
print('possible feature sets', i)
log(['Features %d' % i])
selected_train_X = train_X[possible_feature_sets[i]]
selected_test_X = test_X[possible_feature_sets[i]]
# First we run our non deterministic classifiers a number of times to average their score.
performance_tr_rf = 0
performance_te_rf = 0
for repeat in range(0, N_KCV_REPEATS):
print(datetime.now())
print('\nRepeat', repeat)
print('Random Forest')
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.random_forest(
selected_train_X,
train_y,
selected_test_X,
gridsearch=True,
print_model_details=True)
test_cm = eval.confusion_matrix(test_y, class_test_y,
class_train_prob_y.columns)
DataViz.plot_confusion_matrix(test_cm,
class_train_prob_y.columns,
normalize=False)
performance_tr_rf += eval.accuracy(train_y, class_train_y)
performance_te_rf += eval.accuracy(test_y, class_test_y)
print(datetime.now())
overall_performance_tr_rf = performance_tr_rf / N_KCV_REPEATS
overall_performance_te_rf = performance_te_rf / N_KCV_REPEATS
log([
'RF' + ' train acc: %f' % performance_te_rf +
' test acc: %f' % performance_te_rf
])
# And we run our deterministic classifiers:
print('decision tree')
class_train_y, class_test_y, class_train_prob_y, class_test_prob_y = learner.decision_tree(
selected_train_X,
train_y,
selected_test_X,
gridsearch=True,
print_model_details=True)
performance_tr_dt = eval.accuracy(train_y, class_train_y)
performance_te_dt = eval.accuracy(test_y, class_test_y)
test_cm = eval.confusion_matrix(test_y, class_test_y,
class_train_prob_y.columns)
DataViz.plot_confusion_matrix(test_cm,
class_train_prob_y.columns,
normalize=False)
log([
'DT' + ' train acc: %f' % performance_tr_dt +
' test acc: %f' % performance_te_dt
])
scores_with_sd = util.print_table_row_performances(
feature_names[i], len(selected_train_X.index),
len(selected_test_X.index), [
(overall_performance_tr_rf, overall_performance_te_rf),
(performance_tr_dt, performance_te_dt),
])
scores_over_all_algs.append(scores_with_sd)
DataViz.plot_performances_classification(['RF', 'DT'], feature_names,
scores_over_all_algs)
print(datetime.now())
def main():
# Import the data from the specified location and parse the date index
try:
dataset = pd.read_csv(Path(DATA_PATH / DATASET_FNAME), index_col=0)
dataset.index = pd.to_datetime(dataset.index)
except IOError as e:
print(
'File not found, try to run previous crowdsignals scripts first!')
raise e
# Create an instance of our visualization class to plot the results
DataViz = VisualizeDataset(__file__)
# Compute the number of milliseconds covered by an instance based on the first two rows
milliseconds_per_instance = (dataset.index[1] -
dataset.index[0]).microseconds / 1000
# Create objects for value imputation, low pass filter and PCA
MisVal = ImputationMissingValues()
LowPass = LowPassFilter()
PCA = PrincipalComponentAnalysis()
if FLAGS.mode == 'imputation':
# Impute the missing values and plot an example
imputed_mean_dataset = MisVal.impute_mean(
dataset=copy.deepcopy(dataset), col='hr_watch_rate')
imputed_median_dataset = MisVal.impute_median(
dataset=copy.deepcopy(dataset), col='hr_watch_rate')
imputed_interpolation_dataset = MisVal.impute_interpolate(
dataset=copy.deepcopy(dataset), col='hr_watch_rate')
DataViz.plot_imputed_values(
dataset, ['original', 'mean', 'median', 'interpolation'],
'hr_watch_rate', imputed_mean_dataset['hr_watch_rate'],
imputed_median_dataset['hr_watch_rate'],
imputed_interpolation_dataset['hr_watch_rate'])
elif FLAGS.mode == 'kalman':
# Using the result from Chapter 2, try the Kalman filter on the light_phone_lux attribute and study the result
try:
original_dataset = pd.read_csv(DATA_PATH / ORIG_DATASET_FNAME,
index_col=0)
original_dataset.index = pd.to_datetime(original_dataset.index)
except IOError as e:
print(
'File not found, try to run previous crowdsignals scripts first!'
)
raise e
KalFilter = KalmanFilters()
kalman_dataset = KalFilter.apply_kalman_filter(
data_table=original_dataset, col='acc_phone_x')
DataViz.plot_imputed_values(kalman_dataset, ['original', 'kalman'],
'acc_phone_x',
kalman_dataset['acc_phone_x_kalman'])
DataViz.plot_dataset(data_table=kalman_dataset,
columns=['acc_phone_x', 'acc_phone_x_kalman'],
match=['exact', 'exact'],
display=['line', 'line'])
elif FLAGS.mode == 'lowpass':
# Apply a lowpass filter and reduce the importance of the data above 1.5 Hz
# Determine the sampling frequency
fs = float(1000) / milliseconds_per_instance
# Study acc_phone_x
new_dataset = LowPass.low_pass_filter(
data_table=copy.deepcopy(dataset),
col='acc_phone_x',
sampling_frequency=fs,
cutoff_frequency=FLAGS.cutoff,
order=10)
DataViz.plot_dataset(
new_dataset.iloc[int(0.4 * len(new_dataset.index)
):int(0.43 * len(new_dataset.index)), :],
['acc_phone_x', 'acc_phone_x_lowpass'], ['exact', 'exact'],
['line', 'line'])
elif FLAGS.mode == 'PCA':
# First impute again, as PCA can not deal with missing values
for col in [c for c in dataset.columns if 'label' not in c]:
dataset = MisVal.impute_interpolate(dataset, col)
# Determine the PC's for all but the target columns (the labels and the heart rate)
selected_predictor_cols = [
c for c in dataset.columns
if (not ('label' in c)) and (not (c == 'hr_watch_rate'))
]
pc_values = PCA.determine_pc_explained_variance(
data_table=dataset, cols=selected_predictor_cols)
cumulated_variance = np.cumsum(pc_values)
# Plot the explained variance and cumulated variance
comp_numbers = np.arange(1, len(pc_values) + 1)
DataViz.plot_xy(x=[comp_numbers, comp_numbers],
y=[pc_values, cumulated_variance],
xlabel='principal component number',
ylabel='explained variance',
ylim=[0, 1],
line_styles=['b-', 'r-'],
names=['Variance', 'Cumulated variance'])
# Select 7 as the best number of PC's as this explains most of the variance
n_pcs = 7
dataset = PCA.apply_pca(data_table=copy.deepcopy(dataset),
cols=selected_predictor_cols,
number_comp=n_pcs)
# Visualize the result of the PC's and the overall final dataset
DataViz.plot_dataset(dataset, ['pca_', 'label'], ['like', 'like'],
['line', 'points'])
DataViz.plot_dataset(dataset, [
'acc_', 'gyr_', 'hr_watch_rate', 'light_phone_lux', 'mag_',
'press_phone_', 'pca_', 'label'
], [
'like', 'like', 'like', 'like', 'like', 'like', 'like', 'like',
'like'
], [
'line', 'line', 'line', 'line', 'line', 'line', 'line', 'points',
'points'
])
elif FLAGS.mode == 'final':
# Carry out that operation over all columns except for the label
print('Imputing missing values.')
for col in tqdm([c for c in dataset.columns if 'label' not in c]):
dataset = MisVal.impute_interpolate(dataset=dataset, col=col)
# Include all measurements that have a form of periodicity and filter them
periodic_measurements = [
'acc_phone_x', 'acc_phone_y', 'acc_phone_z', 'acc_watch_x',
'acc_watch_y', 'acc_watch_z', 'gyr_phone_x', 'gyr_phone_y',
'gyr_phone_z', 'gyr_watch_x', 'gyr_watch_y', 'gyr_watch_z',
'mag_phone_x', 'mag_phone_y', 'mag_phone_z', 'mag_watch_x',
'mag_watch_y', 'mag_watch_z'
]
print('Applying low pass filter on peridic measurements.')
# Determine the sampling frequency.
fs = float(1000) / milliseconds_per_instance
for col in tqdm(periodic_measurements):
dataset = LowPass.low_pass_filter(data_table=dataset,
col=col,
sampling_frequency=fs,
cutoff_frequency=FLAGS.cutoff,
order=10)
dataset[col] = dataset[col + '_lowpass']
del dataset[col + '_lowpass']
# Use the optimal found parameter n_pcs = 7 to apply PCA to the final dataset
selected_predictor_cols = [
c for c in dataset.columns
if (not ('label' in c)) and (not (c == 'hr_watch_rate'))
]
n_pcs = 7
dataset = PCA.apply_pca(data_table=copy.deepcopy(dataset),
cols=selected_predictor_cols,
number_comp=n_pcs)
# Visualize the final overall dataset
DataViz.plot_dataset(dataset, [
'acc_', 'gyr_', 'hr_watch_rate', 'light_phone_lux', 'mag_',
'press_phone_', 'pca_', 'label'
], [
'like', 'like', 'like', 'like', 'like', 'like', 'like', 'like',
'like'
], [
'line', 'line', 'line', 'line', 'line', 'line', 'line', 'points',
'points'
])
# Store the final outcome
dataset.to_csv(DATA_PATH / RESULT_FNAME)
## Do some initial runs to determine the right number for k
print('===== kmeans clustering mag =====')
for k in k_values:
print(f'k = {k}')
dataset_cluster = clusteringNH.k_means_over_instances(
copy.deepcopy(dataset), ['gravity.x', 'gravity.y', 'gravity.z'], k,
'default', 20, 10)
silhouette_score = dataset_cluster['silhouette'].mean()
print(f'silhouette = {silhouette_score}')
silhouette_values.append(silhouette_score)
DataViz.plot_xy(x=[k_values],
y=[silhouette_values],
xlabel='k',
ylabel='silhouette score kmeans',
ylim=[0, 1],
line_styles=['b-'])
# And run the knn with the highest silhouette score
# k = 6 # todo: replaced with np.argmax call over silhouette scores
k = k_values[np.argmax(silhouette_values)]
print(f'Highest K-Means silhouette score: k = {k}')
dataset_knn = clusteringNH.k_means_over_instances(
copy.deepcopy(dataset), ['gravity.x', 'gravity.y', 'gravity.z'], k,
'default', 50, 50)
DataViz.plot_clusters_3d(dataset_knn, ['gravity.x', 'gravity.y', 'gravity.z'],
'cluster', ['label'])
DataViz.plot_silhouette(dataset_knn, 'cluster', 'silhouette')