tests_paper+supp.py

import utils_ECAD as AD_algos
import matplotlib.patches as mpatches
import matplotlib.pyplot as pl
from pyod.models.pca import PCA
from pyod.models.ocsvm import OCSVM
from pyod.models.iforest import IForest
from pyod.models.hbos import HBOS
from pyod.models.knn import KNN   # kNN detector
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import GradientBoostingClassifier
from sklearn import neighbors
from sklearn.neural_network import MLPClassifier
from sklearn import svm
from PI_class_EnbPI import prediction_interval
import utils_EnbPI as util
from utils_EnbPI import plot_average_new, grouped_box_new, one_dimen_transform
from matplotlib.lines import Line2D  # For legend handles
import statsmodels as sm
import calendar
import warnings
import matplotlib.pyplot as plt
from sklearn.linear_model import RidgeCV, LassoCV
from sklearn.ensemble import RandomForestRegressor
import itertools
import importlib
import time
import pandas as pd
import numpy as np
import os
import sys
import keras
warnings.filterwarnings("ignore")

'''This File contains code for reproducing all figures
    in the paper (including those in the appendix).
    The main difference is the number of additional experiments being done
    and extra steps needed for network data in California Energy Data'''


'''A. Marginal Coverage Results'''
# Read data and initialize parameters
max_data_size = 10000
# data0--2 results are in Section 8.4
data0 = util.read_data(0, 'Data/green_house_data.csv', max_data_size)
data1 = util.read_data(1, 'Data/appliances_data.csv', max_data_size)
data2 = util.read_data(
    2, 'Data/Beijing_air_Tiantan_data.csv', max_data_size)
dataSolar_Atl = util.read_data(3, 'Data/Solar_Atl_data.csv', max_data_size)
# Results in Sec 8.3
CA_cities = ['Fremont', 'Milpitas', 'Mountain_View', 'North_San_Jose',
             'Palo_Alto', 'Redwood_City', 'San_Mateo', 'Santa_Clara',
             'Sunnyvale']
for city in CA_cities:
    globals()['data%s' % city] = read_CA_data(f'Data/{city}_data.csv')
stride = 1
miss_test_idx = []
alpha = 0.1
tot_trial = 10  # For CP methods that randomizes
np.random.seed(98765)
B = 30  # number of bootstrap samples
Data_name = ['green_house', 'appliances', 'Beijing_air',
             'Solar_Atl', 'Palo_Alto', 'Wind_Austin']
Data_name_network = ['Palo_Alto']
response_ls = {'green_house': 15, 'appliances': 'Appliances',
               'Beijing_air': 'PM2.5', 'Solar_Atl': 'DHI', 'Wind_Austin': 'MWH'}
response_ls_network = {'Palo_Alto': 'DHI'}
data_ind = {}
for i in range(len(Data_name)):
    key = Data_name[i]
    if i <= 2:
        data_ind[key] = i
    else:
        data_ind[key] = key
data_ind_network = {'Palo_Alto': 'Palo_Alto'}
min_alpha = 0.0001
max_alpha = 10
ridge_cv = RidgeCV(alphas=np.linspace(min_alpha, max_alpha, 10))
random_forest = RandomForestRegressor(n_estimators=10, criterion='mse',
                                      bootstrap=False, max_depth=2, n_jobs=-1)


def big_transform(CA_cities, current_city, one_dim, train_size):
    # Next, merge these data (so concatenate X_t and Y_t for one_d or not)
    # Return [X_train, X_test, Y_train, Y_test] from data_x and data_y
    # Data_x is either multivariate (direct concatenation)
    # or univariate (transform each series and THEN concatenate the transformed series)
    big_X_train = []
    big_X_predict = []
    for city in CA_cities:
        data = eval(f'data{city}')  # Pandas DataFrame
        data_x = data.loc[:, data.columns != 'DHI']
        data_y = data['DHI']
        data_x_numpy = data_x.to_numpy()  # Convert to numpy
        data_y_numpy = data_y.to_numpy()  # Convert to numpy
        X_train = data_x_numpy[:train_size, :]
        X_predict = data_x_numpy[train_size:, :]
        Y_train_del = data_y_numpy[:train_size]
        Y_predict_del = data_y_numpy[train_size:]
        if city == current_city:
            Y_train = Y_train_del
            Y_predict = Y_predict_del
        if one_dim:
            X_train, X_predict, Y_train_del, Y_predict_del = one_dimen_transform(
                Y_train_del, Y_predict_del, d=20)
            big_X_train.append(X_train)
            big_X_predict.append(X_predict)
            if city == current_city:
                Y_train = Y_train_del
        else:
            big_X_train.append(X_train)
            big_X_predict.append(X_predict)
    X_train = np.hstack(big_X_train)
    X_predict = np.hstack(big_X_predict)
    return([X_train, X_predict, Y_train, Y_predict])


tot_trial = 10
rnn = False
# True for Palo Alto only, as it is a network. So need to run the procedure TWICE
energy_data = True

''' A.1 Coverage over 1-\alpha (e.g. Figure 1)'''
alpha_ls = np.linspace(0.05, 0.25, 5)
methods = ['Ensemble']
for one_dim in [True, False]:
    for data_name in Data_name:
        data = eval(f'data{data_ind[data_name]}')  # Pandas DataFrame
        data_x = data.loc[:, data.columns != response_ls[data_name]]
        data_y = data[response_ls[data_name]]
        data_x_numpy = data_x.to_numpy()  # Convert to numpy
        data_y_numpy = data_y.to_numpy()  # Convert to numpy
        total_data_points = data_x_numpy.shape[0]
        train_size = int(0.2 * total_data_points)
        results = pd.DataFrame(columns=['itrial', 'dataname', 'muh_fun',
                                        'method', 'alpha', 'coverage', 'width'])
        results_ts = pd.DataFrame(columns=['itrial', 'dataname',
                                           'method', 'alpha', 'coverage', 'width'])
        for itrial in range(tot_trial):
            np.random.seed(98765 + itrial)
            for alpha in alpha_ls:
                # Note, this is necessary because a model may "remember the past"
                nnet = util.keras_mod()
                rnnet = util.keras_rnn()
                print(f'At trial # {itrial} and alpha={alpha}')
                print(f'For {data_name}')
                if energy_data:
                    X_train, X_predict, Y_train, Y_predict = big_transform(
                        Data_name, data_name, one_dim, train_size)
                    d = 20
                else:
                    X_train = data_x_numpy[:train_size, :]
                    X_predict = data_x_numpy[train_size:, :]
                    Y_train = data_y_numpy[:train_size]
                    Y_predict = data_y_numpy[train_size:]
                    d = 20  # for 1-d memory depth
                    if one_dim:
                        X_train, X_predict, Y_train, Y_predict = one_dimen_transform(
                            Y_train, Y_predict, d=d)
                ridge_results = prediction_interval(
                    ridge_cv,  X_train, X_predict, Y_train, Y_predict)
                rf_results = prediction_interval(
                    random_forest,  X_train, X_predict, Y_train, Y_predict)
                nn_results = prediction_interval(
                    nnet,  X_train, X_predict, Y_train, Y_predict)
                if rnn:
                    T, k = X_train.shape
                    T1 = X_predict.shape[0]
                    X_train = X_train.reshape((T, 1, k))
                    X_predict = X_predict.reshape((T1, 1, k))
                    rnn_results = prediction_interval(
                        rnnet, X_train, X_predict, Y_train, Y_predict)
                if itrial == 0:
                    # For ARIMA, only run once
                    result_ts = ridge_results.run_experiments(
                        alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods, none_CP=True)
                    result_ts.rename(
                        columns={'train_size': 'alpha'}, inplace=True)
                    if one_dim:
                        result_ts['alpha'].replace(
                            train_size - d, alpha, inplace=True)
                    else:
                        result_ts['alpha'].replace(
                            train_size, alpha, inplace=True)
                    results_ts = pd.concat([results_ts, result_ts])
                    results_ts.to_csv(
                        f'Results/{data_name}_many_alpha_new_ARIMA.csv', index=False)
                # CP Methods
                print(f'regressor is {ridge_cv.__class__.__name__}')
                result_ridge = ridge_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                print(f'regressor is {random_forest.__class__.__name__}')
                result_rf = rf_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                print(f'regressor is {nnet.name}')
                # start = time.time()
                result_nn = nn_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                if rnn:
                    print(f'regressor is {rnnet.name}')
                    result_rnn = rnn_results.run_experiments(
                        alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                    result_rnn['muh_fun'] = 'RNN'
                    results_now = pd.concat(
                        [result_ridge, result_rf, result_nn, result_rnn])
                else:
                    results_now = pd.concat(
                        [result_ridge, result_rf, result_nn])
                results_now.rename(
                    columns={'train_size': 'alpha'}, inplace=True)
                if one_dim:
                    results_now['alpha'].replace(
                        train_size - d, alpha, inplace=True)
                else:
                    results_now['alpha'].replace(
                        train_size, alpha, inplace=True)
                results = pd.concat([results, results_now])
                if one_dim:
                    results.to_csv(
                        f'Results/{data_name}_many_alpha_new_1d.csv', index=False)
                else:
                    results.to_csv(
                        f'Results/{data_name}_many_alpha_new.csv', index=False)


def merge_arima(data_name, which):
    data1 = pd.read_csv(f'Results/{data_name}_many_alpha_new{which}.csv')
    data2 = pd.read_csv(f'Results/{data_name}_many_alpha_new_ARIMA.csv')
    data1 = pd.concat((data1, data2))
    data1.reset_index(inplace=True)
    print(data1.shape)
    data1.to_csv(f'Results/{data_name}_many_alpha_new{which}.csv', index=False)


for data_name in Data_name:
    merge_arima(data_name, '_1d')
    merge_arima(data_name, '')

''' A.2 Coverage over training size (e.g. Figure 2)'''

if energy_data:
    Data_name = Data_name
    response_ls = response_ls
    data_ind = data_ind
else:
    Data_name = Data_name_network
    response_ls = response_ls_network
    data_ind = data_ind_network
for one_dim in [True, False]:
    methods = ['Ensemble', 'ICP', 'Weighted_ICP']
    # NOTE: if one wants to compare with JaB, replace 'Weighted_ICP' by 'Jab' in methods above.
    # Run Ridge, Lasso, RF, and NN
    for data_name in Data_name:
        data = eval(f'data{data_ind[data_name]}')  # Pandas DataFrame
        data_x = data.loc[:, data.columns != response_ls[data_name]]
        data_y = data[response_ls[data_name]]
        data_x_numpy = data_x.to_numpy()  # Convert to numpy
        data_y_numpy = data_y.to_numpy()  # Convert to numpy
        total_data_points = data_x_numpy.shape[0]
        Train_size = np.linspace(0.1 * total_data_points,
                                 0.3 * total_data_points, 10).astype(int)
        Train_size = [Train_size[0], Train_size[4], Train_size[8]]
        results = pd.DataFrame(columns=['itrial', 'dataname', 'muh_fun',
                                        'method', 'train_size', 'coverage', 'width'])
        for itrial in range(tot_trial):
            np.random.seed(98765 + itrial)
            for train_size in Train_size:
                # Note, this is necessary because a model may "remember the past"
                nnet = util.keras_mod()
                rnnet = util.keras_rnn()
                print(f'At trial # {itrial} and train_size={train_size}')
                print(f'For {data_name}')
                if energy_data:
                    X_train, X_predict, Y_train, Y_predict = big_transform(
                        Data_name, data_name, one_dim, train_size)
                else:
                    X_train = data_x_numpy[:train_size, :]
                    X_predict = data_x_numpy[train_size:, :]
                    Y_train = data_y_numpy[:train_size]
                    Y_predict = data_y_numpy[train_size:]
                    if one_dim:
                        X_train, X_predict, Y_train, Y_predict = one_dimen_transform(
                            Y_train, Y_predict, d=20)
                ridge_results = prediction_interval(
                    ridge_cv, X_train, X_predict, Y_train, Y_predict)
                rf_results = prediction_interval(
                    random_forest,  X_train, X_predict, Y_train, Y_predict)
                nn_results = prediction_interval(
                    nnet,  X_train, X_predict, Y_train, Y_predict)
                if rnn:
                    T, k = X_train.shape
                    T1 = X_predict.shape[0]
                    X_train = X_train.reshape((T, 1, k))
                    X_predict = X_predict.reshape((T1, 1, k))
                    rnn_results = prediction_interval(
                        rnnet, X_train, X_predict, Y_train, Y_predict)
                # For CP Methods
                print(f'regressor is {ridge_cv.__class__.__name__}')
                result_ridge = ridge_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                print(f'regressor is {random_forest.__class__.__name__}')
                result_rf = rf_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                print(f'regressor is {nnet.name}')
                result_nn = nn_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                result_nn['muh_fun'] = 'NeuralNet'
                if rnn:
                    print(f'regressor is {rnnet.name}')
                    result_rnn = rnn_results.run_experiments(
                        alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods)
                    result_rnn['muh_fun'] = 'RNN'
                    results = pd.concat(
                        [results, result_ridge, result_rf, result_nn, result_rnn])
                else:
                    results = pd.concat(
                        [results, result_ridge, result_rf, result_nn])
                if one_dim:
                    results.to_csv(
                        f'Results/{data_name}_many_train_new_1d.csv', index=False)
                else:
                    results.to_csv(
                        f'Results/{data_name}_many_train_new.csv', index=False)

''' Plots'''
Data_name = ['green_house', 'appliances',
             'Beijing_air', 'Solar_Atl', 'Palo_Alto']
'''Mean Coverage over target coverage
    (ideally a straight line by diagonal)'''
alpha_ls = np.linspace(0.05, 0.25, 5)
x_axis = 1 - alpha_ls
x_axis_name = 'alpha'
for dataname in Data_name:
    for two_rows in [False]:
        util.plot_average_new(x_axis, x_axis_name, Dataname=[
                              dataname], two_rows=two_rows)

'''Grouped boxplot, alpha=0.1, over both univariate and multivariate data for
    certain fractions of training data.'''
for data_name in Data_name:
    util.grouped_box_new(data_name, 'coverage')
    util.grouped_box_new(data_name, 'width')
# # NOTE: Run the line below if you have run JaB (see Fig 6 in our paper) and saved results in .csv
# util.grouped_box_new_with_JaB('Solar_Atl')
'''Table 2: Get average coverage, width, and Winkler score over 10 trials into a table'''


def Latex_table_by_regr(array_ls, regr_name, name_ls=[]):
    # For each data_name, two rows (first row LOO, second Ensemble)
    if len(name_ls) > 0:
        names = []
        for name in name_ls:
            names.append([f'{name}',  '', '', ''])
        names = np.hstack(names)
    else:
        return 0
    method = ['ARIMA', 'Ensemble', 'ICP', 'WeightedICP']
    array_t = np.zeros(len(array_ls), dtype=(
        '<U50,<U30, float64, float64, float64'))
    for j in range(len(names)):
        remainder = np.mod(j, 4)
        if remainder == 0:
            if '1d' in regr_name:
                array_t[j] = ' ', ' ', array_ls[j, 0], array_ls[j,
                                                                1], np.round(array_ls[j, 2], 3)
            else:
                name = '\multirow{4}{*}{' + names[j] + '}'
                array_t[j] = name, method[remainder], array_ls[j,
                                                               0], array_ls[j, 1], np.round(array_ls[j, 2], 3)
        else:
            if '1d' in regr_name:
                array_t[j] = ' ', ' ', array_ls[j, 0], array_ls[j,
                                                                1], np.round(array_ls[j, 2], 3)
            else:
                array_t[j] = ' ', method[remainder], array_ls[j,
                                                              0], array_ls[j, 1], np.round(array_ls[j, 2], 3)
    if len(name_ls) > 0:
        np.savetxt(f"cov_wid_score_{regr_name}_solar.txt", array_t, fmt=(
            '%s', '%s', '%1.2f', '%1.2f', '%.2e'), delimiter=' & ', newline=' \\\\\n', comments='')
    else:
        np.savetxt(f"cov_wid_score_{regr_name}.txt", array_t, fmt=(
            '%s', '%s', '%1.2f', '%1.2f', '%.2e'), delimiter=' & ', newline=' \\\\\n', comments='')


# Get average coverage & widths under X% training data as a table
tot_trial = 10
regr_names = {'RidgeCV': 'ridge'}
data_ind = {}
k = 0
for city in CA_cities:
    data_ind[city] = city
    k += 1
rnn = False
Data_name = CA_cities
for one_dim in [True, False]:
    regr_methods_name = ['RidgeCV']
    # For a particular regressor
    methods_name = ['Ensemble', 'ICP', 'Weighted_ICP', 'ARIMA']
    row_len = len(Data_name) * len(methods_name)
    col_len = 3  # coverage, width, Winkler score
    ridge_table_result = np.zeros((row_len, col_len, tot_trial))
    methods = ['Ensemble', 'ICP', 'Weighted_ICP']
    if rnn:
        rnn_table_result = np.zeros((row_len, col_len, tot_trial))
    for itrial in range(tot_trial):
        k = 0
        for data_name in Data_name:
            np.random.seed(98765 + itrial)
            nnet = util.keras_mod()
            rnnet = util.keras_rnn()
            print(f'Trial # {itrial} for data {data_name}')
            data = eval(f'data{data_ind[data_name]}')  # Pandas DataFrame
            data_x = data.loc[:, data.columns != response_ls[data_name]]
            data_y = data[response_ls[data_name]]
            data_x_numpy = data_x.to_numpy()  # Convert to numpy
            data_y_numpy = data_y.to_numpy()  # Convert to numpy
            total_data_points = data_x_numpy.shape[0]
            train_size = int(0.2 * total_data_points)
            results = pd.DataFrame(columns=['itrial', 'dataname', 'muh_fun',
                                            'method', 'train_size', 'coverage', 'width'])
            if energy_data:
                X_train, X_predict, Y_train, Y_predict = big_transform(
                    Data_name, data_name, one_dim, train_size)
            else:
                X_train = data_x_numpy[:train_size, :]
                X_predict = data_x_numpy[train_size:, :]
                Y_train = data_y_numpy[:train_size]
                Y_predict = data_y_numpy[train_size:]
                if one_dim:
                    X_train, X_predict, Y_train, Y_predict = one_dimen_transform(
                        Y_train, Y_predict, d=20)
            if itrial < 1:
                arima_results = prediction_interval(
                    ridge_cv, X_train, X_predict, Y_train, Y_predict)
            ridge_results = prediction_interval(
                ridge_cv, X_train, X_predict, Y_train, Y_predict)
            if itrial < 1:
                PI_cov_wid_arima = arima_results.run_experiments(
                    alpha, B, stride, data_name, itrial, miss_test_idx, get_plots=True, none_CP=True, methods=methods)
            print(f'regressor is Ridge')
            PI_cov_wid_ridge = ridge_results.run_experiments(
                alpha, B, stride, data_name, itrial, miss_test_idx, get_plots=True, methods=methods)
            if itrial < 1:
                arima_scores = arima_results.Winkler_score(
                    PI_cov_wid_arima[:-1][0], Data_name[k], methods, alpha, none_CP=True)[0]
            ridge_scores = ridge_results.Winkler_score(
                PI_cov_wid_ridge[:-1], Data_name[k], methods, alpha)
            # Store results for a particular regressor
            for regr_method in regr_methods_name:
                # Store results
                # Note, *4 because each dataset is coupled with 4 PI methods: ARIMA,EnPI, ICP, and WeightedICP
                regr_name = regr_names[regr_method]
                if itrial < 1:
                    eval(f'{regr_name}_table_result')[k * 4,
                                                      0, :] = PI_cov_wid_arima[-1]['coverage']
                eval(f'{regr_name}_table_result')[1 + k * 4:(k + 1)
                                                  * 4, 0, itrial] = eval(f'PI_cov_wid_{regr_name}')[-1]['coverage']
                eval(f'{regr_name}_table_result')[
                    k * 4, 1, itrial] = PI_cov_wid_arima[-1]['width']
                eval(f'{regr_name}_table_result')[1 + k * 4:(k + 1)
                                                  * 4, 1, itrial] = eval(f'PI_cov_wid_{regr_name}')[-1]['width']
                eval(f'{regr_name}_table_result')[
                    k * 4, 2, itrial] = arima_scores
                eval(f'{regr_name}_table_result')[
                    1 + k * 4:(k + 1) * 4, 2, itrial] = eval(f'{regr_name}_scores')
            print(eval(f'{regr_name}_table_result')[:, :, itrial])
            k += 1
        # Quick check
        np.set_printoptions(precision=2)
        ridge_table_result_save = np.mean(ridge_table_result, axis=2)
        if one_dim:
            # 1d
            Latex_table_by_regr(ridge_table_result_save,
                                'ridge_1d', name_ls=Data_name)
        else:
            # Non-1d
            Latex_table_by_regr(ridge_table_result_save,
                                'ridge', name_ls=Data_name)

'''





'''

'''B. Conditional coverage (e.g. Figure 3)'''
# NOTE: More results/plots than what have been shown in the paper are created
# but the results all look rather similar
# First run the five functions below


def missing_data(data, missing_frac, update=False):
    n = len(data)
    idx = np.random.choice(n, size=int(missing_frac * n), replace=False)
    if update:
        data = np.delete(data, idx, 0)
    idx = idx.tolist()
    return (data, idx)


def restructure_X_t(darray):
    '''
    For each row i after the first row, take i-1 last entries of the first row and then impute the rest
    Imputation is just generating random N(Y_train_mean, Y_train_std), where
    Y_train is the first row.
    '''
    s = darray.shape[1]
    copy = np.copy(darray)
    for i in range(1, min(s, darray.shape[0])):
        copy[i, :s - i] = copy[0, i:]
        imputed_val = np.abs(np.random.normal(loc=np.mean(
            copy[0]), scale=np.std(copy[0]), size=i))
        copy[i, s - i:] = imputed_val
    return copy


def further_preprocess(data, response_name='DHI'):
    '''Extract non-zero hours and also hours between 10AM-2PM (where radiation is high) '''
    max_recorder = pd.DataFrame(np.zeros(24), index=range(0, 24))
    for i in range(0, 24):
        # Check at what times max recording is 0 (meaning no recording yet)
        # 12:00 AM every day. for every later hour, + i \in \{1,...,23\}
        time = np.arange(365) * 24 + i
        max_record = np.max(data[response_name][time])
        max_recorder.iloc[i] = max_record
    # Drop these non-zero things
    data_sub = data.copy()
    to_be_droped = np.where(max_recorder == 0)[0]
    print(to_be_droped)
    drop_idx = []
    if len(to_be_droped) > 0:
        for i in to_be_droped:
            drop_idx.append(np.arange(365) * 24 + i)
        drop_idx = np.hstack(drop_idx)
        data_sub.drop(drop_idx, inplace=True)
    else:
        data_sub = []
    # Create near_noon data between 10AM-2PM
    to_be_included = np.array([10, 11, 12, 13, 14])
    to_be_droped = np.delete(np.arange(24), to_be_included)
    data_near_noon = data.copy()
    drop_idx = []
    for i in to_be_droped:
        drop_idx.append(np.arange(365) * 24 + i)
    drop_idx = np.hstack(drop_idx)
    data_near_noon.drop(drop_idx, inplace=True)
    return [data_sub, data_near_noon]


def big_transform_s_beyond_1(sub, cities, current_city, one_dim, missing, miss_frac=0.25):
    '''Overall, include ALL other cities' data in the CURRENT city being considered.
       1. Check what data is used (full, sub, or near-noon), need sub, but it is now suppressed.
       # NOTE, 1 is suppressed for now, since we are uncertain whether sub or near-noon is needed for Californian results
       2. If missing, process these training and testing data before transform
       -->> Current city and neighbors are assumed to have DIFFERENT missing fractions.
       3. Then, if one_dim, transform data (include past), but since s>1, apply *restructure_X_t* to s rows a time'''
    big_X_train = []
    big_X_predict = []
    for city in cities:
        print(city)
        # Start 1
        data_full = eval(f'data{city}')  # Pandas DataFrame
        if city == 'Wind_Austin':
            data_sub, data_near_noon = further_preprocess(
                data_full, response_name='MWH')
        else:
            data_sub, data_near_noon = further_preprocess(data_full)
        if sub == 0:
            data = data_full
            stride = 24
        elif sub == 1:
            data = data_sub
            stride = int(len(data) / 365)
        else:
            data = data_near_noon
            stride = 5
        train_size = 92 * stride
        col_name = 'MWH' if city == 'Wind_Austin' else 'DHI'
        data_x = data.loc[:, data.columns != col_name]
        data_y = data[col_name]
        data_x_numpy = data_x.to_numpy()  # Convert to numpy
        data_y_numpy = data_y.to_numpy()  # Convert to numpy
        X_train = data_x_numpy[:train_size, :]
        X_predict = data_x_numpy[train_size:, :]
        Y_train_del = data_y_numpy[:train_size]
        Y_predict_del = data_y_numpy[train_size:]
        # Finish 1
        # Start 2
        if missing:
            X_train, miss_train_idx = missing_data(
                X_train, missing_frac=miss_frac, update=True)
            Y_train_del = np.delete(Y_train_del, miss_train_idx)
            Y_predict_del, miss_test_idx = missing_data(
                Y_predict_del, missing_frac=miss_frac, update=False)
            if city == current_city:
                # Need an additional Y_truth
                Y_train = Y_train_del
                Y_predict = Y_predict_del.copy()
                true_miss_text_idx = miss_test_idx
            Y_predict_del[miss_test_idx] = np.abs(np.random.normal(loc=np.mean(
                Y_train_del), scale=np.std(Y_train_del), size=len(miss_test_idx)))

        else:
            true_miss_text_idx = []
            if city == current_city:
                Y_train = Y_train_del
                Y_predict = Y_predict_del
        # Finish 2
        # Start 3
        if one_dim:
            X_train, X_predict, Y_train_del, Y_predict_del = one_dimen_transform(
                Y_train_del, Y_predict_del, d=min(stride, 24))  # Note: this handles 'no_slide (stride=infty)' case
            j = 0
            for k in range(len(X_predict) // stride + 1):
                X_predict[j * k:min((j + 1) * k, len(X_predict))
                          ] = restructure_X_t(X_predict[j * k:min((j + 1) * k, len(X_predict))])
                j += 1
            big_X_train.append(X_train)
            big_X_predict.append(X_predict)
            if city == current_city:
                Y_train = Y_train_del
        else:
            big_X_train.append(X_train)
            big_X_predict.append(X_predict)
        # Finish 3
    X_train = np.abs(np.hstack(big_X_train))
    X_predict = np.abs(np.hstack(big_X_predict))
    return([X_train, X_predict, Y_train, Y_predict, true_miss_text_idx, stride])


def all_together(Data_name, sub, no_slide, missing, miss_frac=0.25, one_dim=False):
    methods = ['Ensemble']
    train_days = 92
    density_est = False
    itrial = 1
    results_ls = {}
    alpha = 0.1
    B = np.random.binomial(100, np.exp(-1))  # number of bootstrap samples
    for data_name in Data_name:
        np.random.seed(98765)
        # nnet = util.keras_mod()  # Note, this is necessary because a model may "remember the past"
        X_train, X_predict, Y_train, Y_predict, miss_test_idx, stride = big_transform_s_beyond_1(
            sub, Data_name, data_name, one_dim, missing)
        train_size = 92 * stride
        print(f'At train_size={train_size}')
        print(f'For {data_name}')
        if no_slide:
            stride = int((365 - 92) * stride)  # No slide at all
        print(stride)
        nnet = util.keras_mod()
        min_alpha = 0.0001
        max_alpha = 10
        ridge_cv = RidgeCV(alphas=np.linspace(min_alpha, max_alpha, 10))
        random_forest = RandomForestRegressor(n_estimators=10, criterion='mse',
                                              bootstrap=False, max_depth=2, n_jobs=-1)
        ridge_results = prediction_interval(
            ridge_cv, X_train, X_predict, Y_train, Y_predict)
        rf_results = prediction_interval(
            random_forest,  X_train, X_predict, Y_train, Y_predict)
        nn_results = prediction_interval(
            nnet,  X_train, X_predict, Y_train, Y_predict)
        # For CP Methods
        print(f'regressor is {ridge_cv.__class__.__name__}')
        result_ridge = ridge_results.run_experiments(
            alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods, get_plots=True)
        print(f'regressor is {random_forest.__class__.__name__}')
        result_rf = rf_results.run_experiments(
            alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods, get_plots=True)
        print(f'regressor is {nnet.name}')
        result_nn = nn_results.run_experiments(
            alpha, B, stride, data_name, itrial, miss_test_idx, methods=methods, get_plots=True)
        results_ls[data_name] = [result_ridge,
                                 result_rf, result_nn, stride, Y_predict]
    return results_ls


'''B.1 Solar ATL (Fig 3 + Sec 8.2) '''
'''1. Multi-step inference: first three are for my EnbPI, last two for no slide (Figure 9)'''
ATL_cities = ['Solar_Atl']
max_data_size = 10000
dataSolar_Atl = util.read_data(
    3, 'Data/Solar_Atl_data.csv', max_data_size)
# s=14, all hours
results_ls_one_d_no_missing_and_slide_sub = all_together(
    Data_name=ATL_cities, sub=1, no_slide=False, missing=False, one_dim=True)
results_ls_no_missing_and_slide_sub = all_together(
    Data_name=ATL_cities, sub=1, no_slide=False, missing=False, one_dim=False)

# s=5, hours near noon
results_ls_one_d_no_missing_and_slide_near_noon = all_together(
    Data_name=ATL_cities, sub=2, no_slide=False, missing=False, one_dim=True)
results_ls_no_missing_and_slide_near_noon = all_together(
    Data_name=ATL_cities, sub=2, no_slide=False, missing=False, one_dim=False)

# No slide: can have poor coverage even if near-noon data is used
results_ls_one_d_no_missing_and_no_slide_near_noon = all_together(
    Data_name=ATL_cities, sub=2, no_slide=True, missing=False, one_dim=True)
results_ls_no_missing_and_no_slide_near_noon = all_together(
    Data_name=ATL_cities, sub=2, no_slide=True, missing=False, one_dim=False)

'''2. Multi-step inference under missing data using EnbPI'''

results_ls_one_d_with_missing_and_slide_sub = all_together(
    Data_name=ATL_cities, sub=1, no_slide=False, missing=True, one_dim=True)
results_ls_with_missing_and_slide_sub = all_together(
    Data_name=ATL_cities, sub=1, no_slide=False, missing=True, one_dim=False)
results_ls_one_d_with_missing_and_slide_near_noon = all_together(
    Data_name=ATL_cities, sub=2, no_slide=False, missing=True, one_dim=True)
results_ls_with_missing_and_slide_near_noon = all_together(
    Data_name=ATL_cities, sub=2, no_slide=False, missing=True, one_dim=False)
'''Plots'''
Data_name = ATL_cities

# s=14
make_cond_plots(Data_name, results_ls_no_missing_and_slide_sub,
                no_slide=False, missing=False, one_d=False, five_in_a_row=False)
make_cond_plots(Data_name, results_ls_one_d_no_missing_and_slide_sub,
                no_slide=False, missing=False, one_d=True, five_in_a_row=False)
make_cond_plots(Data_name, results_ls_with_missing_and_slide_sub,
                no_slide=False, missing=True, one_d=False, five_in_a_row=False)
make_cond_plots(Data_name, results_ls_one_d_with_missing_and_slide_sub,
                no_slide=False, missing=True, one_d=True, five_in_a_row=False)

# s=5

make_cond_plots(Data_name, results_ls_one_d_no_missing_and_slide_near_noon,
                no_slide=False, missing=False, one_d=True)
make_cond_plots(Data_name, results_ls_no_missing_and_slide_near_noon,
                no_slide=False, missing=False, one_d=False)
make_cond_plots(Data_name, results_ls_one_d_no_missing_and_no_slide_near_noon,
                no_slide=True, missing=False, one_d=True)
make_cond_plots(Data_name, results_ls_no_missing_and_no_slide_near_noon,
                no_slide=True, missing=False, one_d=False)
make_cond_plots(Data_name, results_ls_one_d_with_missing_and_slide_near_noon,
                no_slide=False, missing=True, one_d=True)
make_cond_plots(Data_name, results_ls_with_missing_and_slide_near_noon,
                no_slide=False, missing=True, one_d=False)
'''




'''

'''B.2 Solar CA (Sec 8.3)'''
CA_cities = ['Fremont', 'Milpitas', 'Mountain_View', 'North_San_Jose',
             'Palo_Alto', 'Redwood_City', 'San_Mateo', 'Santa_Clara',
             'Sunnyvale']
CA_cities = ['Palo_Alto']
for city in CA_cities:
    globals()['data%s' % city] = read_CA_data(f'Data/{city}_data.csv')

'''1. Multi-step inference: first two are for my EnbPI, last two for no slide'''
# Returns a dictionary with results at each city in CA
# s=15 (not 14)
results_ls_one_d_no_missing_and_slide_sub = all_together(
    Data_name=CA_cities, sub=1, no_slide=False, missing=False, one_dim=True)
results_ls_no_missing_and_slide_sub = all_together(
    Data_name=CA_cities, sub=1, no_slide=False, missing=False, one_dim=False)


# s=5
results_ls_one_d_no_missing_and_slide_near_noon = all_together(
    Data_name=CA_cities, sub=2, no_slide=False, missing=False, one_dim=True)
results_ls_no_missing_and_slide_near_noon = all_together(
    Data_name=CA_cities, sub=2, no_slide=False, missing=False, one_dim=False)

# No_slide: can have poor coverage even if near-noon data is used
results_ls_one_d_no_missing_and_no_slide_near_noon = all_together(
    Data_name=CA_cities, sub=2, no_slide=True, missing=False, one_dim=True)

results_ls_no_missing_and_no_slide_near_noon = all_together(
    Data_name=CA_cities, sub=2, no_slide=True, missing=False, one_dim=False)

'''2. Multi-step inference under missing data using EnbPI'''
# s=15 (not 14)
results_ls_one_d_with_missing_and_slide_sub = all_together(
    Data_name=CA_cities, sub=1, no_slide=False, missing=True, one_dim=True)
results_ls_with_missing_and_slide_sub = all_together(
    Data_name=CA_cities, sub=1, no_slide=False, missing=True, one_dim=False)

# s=5

results_ls_one_d_with_missing_and_slide_near_noon = all_together(
    Data_name=CA_cities, sub=2, no_slide=False, missing=True, one_dim=True)

results_ls_with_missing_and_slide_near_noon = all_together(
    Data_name=CA_cities, sub=2, no_slide=False, missing=True, one_dim=False)

'''Plots'''

Data_name = CA_cities
# s=14
make_cond_plots(Data_name, results_ls_no_missing_and_slide_sub,
                no_slide=False, missing=False, one_d=False, five_in_a_row=False)
make_cond_plots(Data_name, results_ls_one_d_no_missing_and_slide_sub,
                no_slide=False, missing=False, one_d=True, five_in_a_row=False)
make_cond_plots(Data_name, results_ls_with_missing_and_slide_sub,
                no_slide=False, missing=True, one_d=False, five_in_a_row=False)
make_cond_plots(Data_name, results_ls_one_d_with_missing_and_slide_sub,
                no_slide=False, missing=True, one_d=True, five_in_a_row=False)

# s=5

make_cond_plots(Data_name, results_ls_one_d_no_missing_and_slide_near_noon,
                no_slide=False, missing=False, one_d=True)
make_cond_plots(Data_name, results_ls_no_missing_and_slide_near_noon,
                no_slide=False, missing=False, one_d=False)
make_cond_plots(Data_name, results_ls_one_d_no_missing_and_no_slide_near_noon,
                no_slide=True, missing=False, one_d=True)
make_cond_plots(Data_name, results_ls_no_missing_and_no_slide_near_noon,
                no_slide=True, missing=False, one_d=False)
make_cond_plots(Data_name, results_ls_one_d_with_missing_and_slide_near_noon,
                no_slide=False, missing=True, one_d=True)
make_cond_plots(Data_name, results_ls_with_missing_and_slide_near_noon,
                no_slide=False, missing=True, one_d=False)


'''


'''

'''B.3 Wind (Sec 8.3)
   From https://github.com/Duvey314/austin-green-energy-predictor'''
Wind_cities = ['Wind_Austin']
dataWind_Austin = read_wind_data()
# Note, no missing entries, so only sub=0 used.
results_ls_one_d_no_missing_and_slide_full = all_together(
    Data_name=Wind_cities, sub=0, no_slide=False, missing=False, one_dim=True)

results_ls_one_d_with_missing_and_slide_full = all_together(
    Data_name=Wind_cities, sub=0, no_slide=False, missing=True, one_dim=True)

'''Plots'''

Data_name = Wind_cities

make_cond_plots(Data_name, results_ls_one_d_no_missing_and_slide_full,
                no_slide=False, missing=False, one_d=True)
make_cond_plots(Data_name, results_ls_one_d_with_missing_and_slide_full,
                no_slide=False, missing=True, one_d=True)

'''






'''

'''Figure 4: Anomaly Detection by ECAD '''
'''Credit Card Fraud # 2: Data retrieved here: https://www.kaggle.com/mlg-ulb/creditcardfraud
    Data shape is (284807, 31)'''
# HBOS: https://pyod.readthedocs.io/en/latest/pyod.models.html#module-pyod.models.hbos
# IForest: https://pyod.readthedocs.io/en/latest/pyod.models.html#module-pyod.models.iforest
# OCSVM https://pyod.readthedocs.io/en/latest/pyod.models.html#module-pyod.models.ocsvm
# PCA: https://pyod.readthedocs.io/en/latest/pyod.models.html#module-pyod.models.pca

'''Get data and downsample'''


def get_data_and_true_abnormal(tot_len):
    dataset = pd.read_csv('Data/Money_Laundry.csv', nrows=tot_len)
    dataset.drop('Time', axis=1, inplace=True)
    true_abnormal = np.where(dataset.Class == 1)[0]
    return dataset, true_abnormal


def get_downsample_data(data, train_size):
    data_x = data.iloc[:, :-1]
    data_y = data.iloc[:, -1]
    data_x_numpy = data_x.to_numpy()  # Convert to numpy
    data_y_numpy = data_y.to_numpy()  # Convert to numpy
    X_train = data_x_numpy[:train_size, :]
    X_predict = data_x_numpy[train_size:, :]
    Y_train = data_y_numpy[:train_size]
    Y_predict = data_y_numpy[train_size:]
    train_abnormal = np.where(Y_train == 1)[0]
    train_normal = np.where(Y_train == 0)[0]
    X_train_abnormal = X_train[train_abnormal, :]
    Y_train_abnormal = Y_train[train_abnormal]
    down_sample_idx = np.random.choice(
        train_normal, 5 * len(train_abnormal), replace=False)
    X_train_normal = X_train[down_sample_idx, :]
    Y_train_normal = Y_train[down_sample_idx]
    X_train = np.vstack((X_train_abnormal, X_train_normal))
    Y_train = np.hstack((Y_train_abnormal, Y_train_normal))
    return (X_train, X_predict, Y_train, Y_predict)


'''(For final results) Put everything together (with competing methods)'''

'''First define functions'''
# Competing methods


def mod_to_result(regr_name, X_train, Y_train, test_true_abnormal):
    mod = eval(regr_name)
    mod.fit(X_train, Y_train)
    est_anomalies = mod.predict(X_predict)
    est_anomalies = np.where(est_anomalies == 1)[0]
    precision, recall, F1 = AD_algos.accuracies(
        est_anomalies, test_true_abnormal)
    return [precision, recall, F1]


def ECAD(tot_size, train_frac):
    data, true_abnormal = get_data_and_true_abnormal(tot_size)
    data.shape
    train_size = int(data.shape[0] * train_frac)
    train_size  # A lot of anomalies occurred around 6000
    neighbor_size = 5  # for each abnormal idx in training, how many of its neighbors are used in calibrating residuals
    # @ 0.005, ~80/60/70. @ 0.1, almost exactly the same
    alpha = np.linspace(0.05, 0.15, 3)
    alpha = [0.05]
    dotted = True
    # A large stride, suitable when training data is large (so less percentile needed)
    stride = 1
    # NOTE: actual getting residual part is the most expensive (since iterate through n...)
    # Thus, let n be even smaller
    return_fitted = False
    est_anomalies = AD_algos.get_anomalies_classification(
        data, 'RF', train_size, alpha,  stride, dotted, return_fitted=return_fitted, neighbor_size=neighbor_size)
    if len(alpha) > 1:
        AD_algos.PR_curve(
            est_anomalies, true_abnormal[true_abnormal > train_size] - train_size)
    else:
        precision, recall, F1 = AD_algos.accuracies(
            est_anomalies[0], true_abnormal[true_abnormal > train_size] - train_size)
        return [precision, recall, F1]


'''Next experiments'''
tot_trial = 1
train_fracs = np.linspace(0.3, 0.7, 5)
tot_size = 284807
methods = ['ECAD', 'HBOS()', 'IForest()', 'OCSVM()', 'PCA()', 'svm.SVC(gamma="auto")', 'GradientBoostingClassifier(n_estimators=100, learning_rate=1.0,max_depth=1, random_state=0)',
           'neighbors.KNeighborsClassifier(n_neighbors=20, weights="distance")', 'MLPClassifier(solver="lbfgs", alpha=1e-5,hidden_layer_sizes=(5, 2), random_state=1)']
method_name = {'ECAD': 'ECAD', 'HBOS()': 'HBOS', 'IForest()': 'IForest', 'OCSVM()': "OCSVM", 'PCA()': 'PCA', 'svm.SVC(gamma="auto")': 'SVC', 'GradientBoostingClassifier(n_estimators=100, learning_rate=1.0,max_depth=1, random_state=0)': 'GBoosting',
               'neighbors.KNeighborsClassifier(n_neighbors=20, weights="distance")': 'KNN', 'MLPClassifier(solver="lbfgs", alpha=1e-5,hidden_layer_sizes=(5, 2), random_state=1)': 'MLPClassifer'}
results = pd.DataFrame(columns=['itrial', 'train_frac', 'method',
                                'precision', 'recall', 'F1'])
for itrial in range(tot_trial):
    np.random.seed(98765 + itrial)
    for train_frac in train_fracs:
        train_frac = np.round(train_frac, 2)
        data, true_abnormal = get_data_and_true_abnormal(tot_size)
        train_size = int(tot_size * train_frac)
        X_train, X_predict, Y_train, _ = get_downsample_data(data, train_size)
        test_true_abnormal = true_abnormal[true_abnormal >
                                           train_size] - train_size
        for method in methods:
            if method == 'ECAD':
                precision, recall, F1 = ECAD(tot_size, train_frac)
            else:
                precision, recall, F1 = mod_to_result(
                    method, X_train, Y_train, test_true_abnormal)
            results.loc[len(results)] = [itrial, train_frac,
                                         method_name[method], precision, recall, F1]
            results.to_csv(f'Results/Kaggle_results.csv', index=False)


results = pd.read_csv('Results/Kaggle_results.csv')
AD_algos.plt_prec_recall_F1(results)