GCN Algorithm Example 1

ST-GCN
GNAR
Author

SEOYEON CHOI

Published

January 11, 2023

Our method; GNAR Dataset Example(fiveVTS, fiveNet)

import

import rpy2
import rpy2.robjects as ro 
from rpy2.robjects.vectors import FloatVector 
from rpy2.robjects.packages import importr

import torch
import numpy as np
from tqdm import tqdm

import torch.nn.functional as F
from torch_geometric_temporal.nn.recurrent import GConvGRU

import matplotlib.pyplot as plt
import pandas as pd

import time

from scipy.interpolate import interp1d
class RecurrentGCN(torch.nn.Module):
    def __init__(self, node_features, filters):
        super(RecurrentGCN, self).__init__()
        self.recurrent = GConvGRU(node_features, filters, 2)
        self.linear = torch.nn.Linear(filters, 1)

    def forward(self, x, edge_index, edge_weight):
        h = self.recurrent(x, edge_index, edge_weight)
        h = F.relu(h)
        h = self.linear(h)
        return h

R

%load_ext rpy2.ipython
%%R
library(GNAR)
library(igraph)

시간측정 방법

t1= time.time()
for epoc in range(1000):
    ## 1 
    yhat = net(x) 
    ## 2 
    loss = loss_fn(yhat,y)
    ## 3 
    loss.backward()
    ## 4 
    optimizr.step()
    optimizr.zero_grad()
t2 = time.time()
t2-t1

EbayesThresh

%%R
library(EbayesThresh)
set.seed(1)
epsilon = rnorm(1000)
signal = sample(c(runif(25,-2,-1.5), runif(25,1.5,2), rep(0,950)))
index_of_trueoutlier = which(signal!=0)
x_ebayes=signal+epsilon
%R -o x_ebayes
%R -o index_of_trueoutlier
%R -o signal
%R -i 
arr = np.array([1,2,3])
%R -i arr
%%R
arr
[1] 1 2 3
ebayesthresh = importr('EbayesThresh').ebayesthresh
outlier_true_index = index_of_trueoutlier
outlier_true_value = x_ebayes[index_of_trueoutlier]
outlier_true_one = signal.copy()
outlier_true_one = list(map(lambda x: -1 if x!=0 else 1,outlier_true_one))

0. 데이터 가정

데이터가정: missing (결측값) 이 있는 자료를 가정

  • missing이 있는상태: ST-GCN 을 사용할 수 X \(\to\) 코드에 에러는 나지 않지만 yhat이 산출되지 않았습니다.
  • missing이 없어야만: ST-GCN 을 사용할 수 O

예제 1 GNAR fiveVT

증명 - missing이 있는상태: ST-GCN 을 사용할 수 X

%%R
summary(fiveNet)
GNARnet with 5 nodes and 10 edges
 of equal length  1
%%R
edges <- as.matrix(fiveNet)
data("fiveNode")
%R -o fiveVTS
%R -o edges
  • node: 5
  • time 200
edges_tensor = torch.tensor(edges)
nonzero_indices = edges_tensor.nonzero()
fiveNet_edge = np.array(nonzero_indices).T

데이터 일부 missing 처리

1) Block 처리

[1] ST-GCN

%%R
fiveVTS0 <- fiveVTS
fiveVTS0[50:150, 3] <- NA
plt.plot(fiveVTS0[:,2])

T = 200
N = 5 # number of Nodes
E = fiveNet_edge
V = np.array([1,2,3,4,5])
t = np.arange(0,T)
node_features = 1
f = torch.tensor(fiveVTS0).reshape(200,5,1).float()
X = f[:199,:,:]
y = f[1:,:,:]
edge_index = torch.tensor(E)
edge_attr = torch.tensor(np.array([1,1,1,1,1,1,1,1,1,1]),dtype=torch.float32)
_ee = enumerate(zip(X,y))
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X,y)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.53it/s]
yhat = torch.stack([model(xt, edge_index, edge_attr) for xt in X]).detach().numpy()
plt.plot(yhat[:,0].data)
plt.plot(yhat[:,1].data)
plt.plot(yhat[:,2].data)
plt.plot(yhat[:,3].data)

2) Random missing values

%%R
set.seed(1)
fiveVTSrandom <- fiveVTS
sampleindex = sort(sample(1:200, 100))
fiveVTSrandom[sampleindex,3] <- NA
%R -o fiveVTSrandom
%R -o sampleindex
plt.plot(fiveVTSrandom[:,2],'o')

3) By 2

%%R
fiveVTStwo <- fiveVTS
indextwo <- rep(seq(1, by = 2, 200))
fiveVTStwo[indextwo, 3] <- NA
%R -o fiveVTStwo
%R -o indextwo
plt.plot(fiveVTStwo[:,2],'o')

1. missing을 채움. (mean, linear interpolation)

1.1. Mean

1) Block

fiveVTS0_mean = fiveVTS0.copy()
fiveVTS0_mean[49:150,2] = np.mean(fiveVTS0[:49,2].tolist()+fiveVTS0[150:,2].tolist())
plt.plot(fiveVTS0_mean[:,2])

2) Random missing values

fiveVTSrandom_mean = fiveVTSrandom.copy()
df = pd.DataFrame(fiveVTSrandom[:,2])
mean_value = df.mean() # finds the mean value of the column A
df = df.fillna(mean_value) # replace missing values with the mean value
fiveVTSrandom_mean[:,2] = np.array(df).reshape(200,)
plt.plot(fiveVTSrandom_mean[:,2])

3) By 2

fiveVTStwo_mean = fiveVTStwo.copy()
df = pd.DataFrame(fiveVTStwo[:,2])
mean_value = df.mean() # finds the mean value of the column A
df = df.fillna(mean_value) # replace missing values with the mean value
fiveVTStwo_mean[:,2] = np.array(df).reshape(200,)
plt.plot(fiveVTStwo_mean[:,2])

1.2. linear interpolation

1) Block

fiveVTS0_linearinterpolation = fiveVTS0.copy()
# Sample data points
x = np.array([48,150])
y = np.array([fiveVTS0_linearinterpolation[48,2],fiveVTS0_linearinterpolation[150,2]])

# Create interpolating function
f = interp1d(x, y, kind='linear')

# Estimate y value for x = 2.5
y_interp = f(range(49,150))
fiveVTS0_linearinterpolation[49:150,2] = y_interp
plt.plot(fiveVTS0_linearinterpolation[:,2])

2) Random missing values

fiveVTSrandom_linearinterpolation = fiveVTSrandom.copy()
_df = pd.DataFrame(fiveVTSrandom_linearinterpolation[:,2])
_df.interpolate(method='linear', inplace=True)
_df = _df.fillna(0)
fiveVTSrandom_linearinterpolation[:,2] = np.array(_df).reshape(200,)
plt.plot(fiveVTSrandom_linearinterpolation[:,2])

3) By 2

fiveVTStwo_linearinterpolation = fiveVTStwo.copy()
_df = pd.Series(fiveVTStwo_linearinterpolation[:,2])
_df.interpolate(method='linear', inplace=True)
_df = _df.fillna(0)
fiveVTStwo_linearinterpolation[:,2] = _df
plt.plot(fiveVTStwo_linearinterpolation[:,2])

2. ST-GCN 을 사용하여 fhat을 구함. (스무딩1)

T = 200
N = 5 # number of Nodes
E = fiveNet_edge
V = np.array([1,2,3,4,5])
t = np.arange(0,T)
node_features = 1
edge_index = torch.tensor(E)
edge_attr = torch.tensor(np.array([1,1,1,1,1,1,1,1,1,1]),dtype=torch.float32)

GNAR이랑 비교 - MSE 등 평가지표

2.1. Mean

1) Block

f_mean = torch.tensor(fiveVTS0_mean).reshape(200,5,1).float()
X_mean = f_mean[:199,:,:]
y_mean = f_mean[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X_mean,y_mean)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.56it/s]
fhat_mean = torch.stack([model(xt, edge_index, edge_attr) for xt in X_mean]).detach().numpy()
plt.plot(fhat_mean[:,2].data)

2) Random missing values

f_fiveVTSrandom_mean = torch.tensor(fiveVTSrandom_mean).reshape(200,5,1).float()
X_fiveVTSrandom_mean = f_fiveVTSrandom_mean[:199,:,:]
y_fiveVTSrandom_mean = f_fiveVTSrandom_mean[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X_fiveVTSrandom_mean,y_fiveVTSrandom_mean)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.55it/s]
fhat_fiveVTSrandom_mean = torch.stack([model(xt, edge_index, edge_attr) for xt in X_fiveVTSrandom_mean]).detach().numpy()
plt.plot(fhat_fiveVTSrandom_mean[:,2].data)

3) By 2

f_fiveVTStwo_mean = torch.tensor(fiveVTStwo_mean).reshape(200,5,1).float()
X_fiveVTStwo_mean = f_fiveVTStwo_mean[:199,:,:]
y_fiveVTStwo_mean = f_fiveVTStwo_mean[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X_fiveVTStwo_mean,y_fiveVTStwo_mean)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.54it/s]
fhat_fiveVTStwo_mean = torch.stack([model(xt, edge_index, edge_attr) for xt in X_fiveVTStwo_mean]).detach().numpy()
plt.plot(fhat_fiveVTStwo_mean[:,2].data)

2.2. linear interpolation

1) Block

f_linearinterpolation = torch.tensor(fiveVTS0_linearinterpolation).reshape(200,5,1).float()
X_linearinterpolation = f_linearinterpolation[:199,:,:]
y_linearinterpolation = f_linearinterpolation[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X_linearinterpolation,y_linearinterpolation)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.55it/s]
fhat_linearinterpolation = torch.stack([model(xt, edge_index, edge_attr) for xt in X_linearinterpolation]).detach().numpy()
plt.plot(fhat_linearinterpolation[:,2].data)

2) Random missing values

f_fiveVTSrandom_linearinterpolation = torch.tensor(fiveVTSrandom_linearinterpolation).reshape(200,5,1).float()
X_fiveVTSrandom_linearinterpolation = f_fiveVTSrandom_linearinterpolation[:199,:,:]
y_fiveVTSrandom_linearinterpolation = f_fiveVTSrandom_linearinterpolation[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X_fiveVTSrandom_linearinterpolation,y_fiveVTSrandom_linearinterpolation)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.55it/s]
fhat_fiveVTSrandom_linearinterpolation = torch.stack([model(xt, edge_index, edge_attr) for xt in X_fiveVTSrandom_linearinterpolation]).detach().numpy()
plt.plot(fhat_fiveVTSrandom_linearinterpolation[:,2].data)

3) By 2

f_fiveVTStwo_linearinterpolation = torch.tensor(fiveVTStwo_linearinterpolation).reshape(200,5,1).float()
X_fiveVTStwo_linearinterpolation = f_fiveVTSrandom_linearinterpolation[:199,:,:]
y_fiveVTStwo_linearinterpolation = f_fiveVTSrandom_linearinterpolation[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X_fiveVTStwo_linearinterpolation,y_fiveVTStwo_linearinterpolation)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.55it/s]
fhat_fiveVTStwo_linearinterpolation = torch.stack([model(xt, edge_index, edge_attr) for xt in X_fiveVTStwo_linearinterpolation]).detach().numpy()
plt.plot(fhat_fiveVTStwo_linearinterpolation[:,2].data)

2.3. 원래 f

f = torch.tensor(fiveVTS).reshape(200,5,1).float()
X = f[:199,:,:]
y = f[1:,:,:]
model = RecurrentGCN(node_features=1, filters=4)

optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

model.train()

for epoch in tqdm(range(50)):
    for time, (xt,yt) in enumerate(zip(X,y)):
        y_hat = model(xt, edge_index, edge_attr)
        cost = torch.mean((y_hat-yt)**2)
        cost.backward()
        optimizer.step()
        optimizer.zero_grad()
100%|██████████| 50/50 [00:32<00:00,  1.55it/s]
fhat_fiveVTS = torch.stack([model(xt, edge_index, edge_attr) for xt in X]).detach().numpy()
plt.plot(fhat_fiveVTS[:,2].data)

3. 2에서 얻은 fhat을 이용하여 그래프퓨리에변환+Ebayesthresh (스무딩2)

  • 그래프퓨리에변환을 하는 가중치
  • 년도끼리 이어주어서 하나의 큰 그래프로 만든뒤에 GFT
  • Ebayesthresh

3.1. Mean

3.1.1. Temporal

1) Block

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])    
fhatbar = np.hstack([Psi[i] @ fhat_mean[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_mean_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_mean_temporal[:,0])
plt.plot(fhatbarhat_mean_temporal[:,1])
plt.plot(fhatbarhat_mean_temporal[:,2])
plt.plot(fhatbarhat_mean_temporal[:,3])
plt.plot(fhatbarhat_mean_temporal[:,4])

2) Random missing values

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])    
fhatbar = np.hstack([Psi[i] @ fhat_fiveVTSrandom_mean[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_random_mean_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_random_mean_temporal[:,0])
plt.plot(fhatbarhat_random_mean_temporal[:,1])
plt.plot(fhatbarhat_random_mean_temporal[:,2])
plt.plot(fhatbarhat_random_mean_temporal[:,3])
plt.plot(fhatbarhat_random_mean_temporal[:,4])

3) By 2

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])    
fhatbar = np.hstack([Psi[i] @ fhat_fiveVTStwo_mean[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_twomean_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_twomean_temporal[:,0])
plt.plot(fhatbarhat_twomean_temporal[:,1])
plt.plot(fhatbarhat_twomean_temporal[:,2])
plt.plot(fhatbarhat_twomean_temporal[:,3])
plt.plot(fhatbarhat_twomean_temporal[:,4])

3.1.2. Spatio

1) Block

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_mean.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_mean_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_mean_spatio[:,0])
plt.plot(fhatbarhat_mean_spatio[:,1])
plt.plot(fhatbarhat_mean_spatio[:,2])
plt.plot(fhatbarhat_mean_spatio[:,3])
plt.plot(fhatbarhat_mean_spatio[:,4])

2) Random missing values

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTSrandom_mean.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_random_mean_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_random_mean_spatio[:,0])
plt.plot(fhatbarhat_random_mean_spatio[:,1])
plt.plot(fhatbarhat_random_mean_spatio[:,2])
plt.plot(fhatbarhat_random_mean_spatio[:,3])
plt.plot(fhatbarhat_random_mean_spatio[:,4])

3) By 2

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTStwo_mean.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_two_mean_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_two_mean_spatio[:,0])
plt.plot(fhatbarhat_two_mean_spatio[:,1])
plt.plot(fhatbarhat_two_mean_spatio[:,2])
plt.plot(fhatbarhat_two_mean_spatio[:,3])
plt.plot(fhatbarhat_two_mean_spatio[:,4])

3.1.3. Spatio-Temporal

1) Block

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_mean.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_mean_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhatbarhat_mean_spatio_temporal[:,0])
plt.plot(fhatbarhat_mean_spatio_temporal[:,1])
plt.plot(fhatbarhat_mean_spatio_temporal[:,2])
plt.plot(fhatbarhat_mean_spatio_temporal[:,3])
plt.plot(fhatbarhat_mean_spatio_temporal[:,4])

2) Random missing values

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTSrandom_mean.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_random_mean_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhatbarhat_random_mean_spatio_temporal[:,0])
plt.plot(fhatbarhat_random_mean_spatio_temporal[:,1])
plt.plot(fhatbarhat_random_mean_spatio_temporal[:,2])
plt.plot(fhatbarhat_random_mean_spatio_temporal[:,3])
plt.plot(fhatbarhat_random_mean_spatio_temporal[:,4])

3) By 2

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTStwo_mean.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_two_mean_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhatbarhat_two_mean_spatio_temporal[:,0])
plt.plot(fhatbarhat_two_mean_spatio_temporal[:,1])
plt.plot(fhatbarhat_two_mean_spatio_temporal[:,2])
plt.plot(fhatbarhat_two_mean_spatio_temporal[:,3])
plt.plot(fhatbarhat_two_mean_spatio_temporal[:,4])

3.2.linear interpolation

3.2.1. Temporal

1) Block

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])
fhatbar = np.hstack([Psi[i] @ fhat_linearinterpolation[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_linearinterpolation_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_linearinterpolation_temporal[:,0])
plt.plot(fhatbarhat_linearinterpolation_temporal[:,1])
plt.plot(fhatbarhat_linearinterpolation_temporal[:,2])
plt.plot(fhatbarhat_linearinterpolation_temporal[:,3])
plt.plot(fhatbarhat_linearinterpolation_temporal[:,4])

2) Random missing values

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])
fhatbar = np.hstack([Psi[i] @ fhat_fiveVTSrandom_linearinterpolation[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_random_linearinterpolation_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_random_linearinterpolation_temporal[:,0])
plt.plot(fhatbarhat_random_linearinterpolation_temporal[:,1])
plt.plot(fhatbarhat_random_linearinterpolation_temporal[:,2])
plt.plot(fhatbarhat_random_linearinterpolation_temporal[:,3])
plt.plot(fhatbarhat_random_linearinterpolation_temporal[:,4])

3) By 2

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])
fhatbar = np.hstack([Psi[i] @ fhat_fiveVTStwo_linearinterpolation[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_two_linearinterpolation_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_two_linearinterpolation_temporal[:,0])
plt.plot(fhatbarhat_two_linearinterpolation_temporal[:,1])
plt.plot(fhatbarhat_two_linearinterpolation_temporal[:,2])
plt.plot(fhatbarhat_two_linearinterpolation_temporal[:,3])
plt.plot(fhatbarhat_two_linearinterpolation_temporal[:,4])

3.2.2. Spatio

1) Block

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_linearinterpolation.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_linearinterpolation_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_linearinterpolation_spatio[:,0])
plt.plot(fhatbarhat_linearinterpolation_spatio[:,1])
plt.plot(fhatbarhat_linearinterpolation_spatio[:,2])
plt.plot(fhatbarhat_linearinterpolation_spatio[:,3])
plt.plot(fhatbarhat_linearinterpolation_spatio[:,4])

2) Random missing values

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTSrandom_linearinterpolation.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_random_linearinterpolation_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_random_linearinterpolation_spatio[:,0])
plt.plot(fhatbarhat_random_linearinterpolation_spatio[:,1])
plt.plot(fhatbarhat_random_linearinterpolation_spatio[:,2])
plt.plot(fhatbarhat_random_linearinterpolation_spatio[:,3])
plt.plot(fhatbarhat_random_linearinterpolation_spatio[:,4])

3) By 2

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTStwo_linearinterpolation.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_two_linearinterpolation_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_two_linearinterpolation_spatio[:,0])
plt.plot(fhatbarhat_two_linearinterpolation_spatio[:,1])
plt.plot(fhatbarhat_two_linearinterpolation_spatio[:,2])
plt.plot(fhatbarhat_two_linearinterpolation_spatio[:,3])
plt.plot(fhatbarhat_two_linearinterpolation_spatio[:,4])

3.2.3. Spatio-Temporal

1) Block

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_linearinterpolation.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhat_linearinterpolation_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhat_linearinterpolation_spatio_temporal[:,0])
plt.plot(fhat_linearinterpolation_spatio_temporal[:,1])
plt.plot(fhat_linearinterpolation_spatio_temporal[:,2])
plt.plot(fhat_linearinterpolation_spatio_temporal[:,3])
plt.plot(fhat_linearinterpolation_spatio_temporal[:,4])

2) Random missing values

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTSrandom_linearinterpolation.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhat_random_linearinterpolation_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhat_random_linearinterpolation_spatio_temporal[:,0])
plt.plot(fhat_random_linearinterpolation_spatio_temporal[:,1])
plt.plot(fhat_random_linearinterpolation_spatio_temporal[:,2])
plt.plot(fhat_random_linearinterpolation_spatio_temporal[:,3])
plt.plot(fhat_random_linearinterpolation_spatio_temporal[:,4])

3) By 2

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTStwo_linearinterpolation.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhat_two_linearinterpolation_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhat_two_linearinterpolation_spatio_temporal[:,0])
plt.plot(fhat_two_linearinterpolation_spatio_temporal[:,1])
plt.plot(fhat_two_linearinterpolation_spatio_temporal[:,2])
plt.plot(fhat_two_linearinterpolation_spatio_temporal[:,3])
plt.plot(fhat_two_linearinterpolation_spatio_temporal[:,4])

3.3. original

1) Temporal

w=np.zeros((5,199,199))
for k in range(5):
    for i in range(199):
        for j in range(199):
            if i==j :
                w[k,i,j] = 0
            elif np.abs(i-j) <= 1 : 
                w[k,i,j] = 1
d = np.array([w[i].sum(axis=1) for i in range(5)])
D= np.array([np.diag(d[i]) for i in range(5)])
L = np.array([np.diag(1/np.sqrt(d[i])) @ (D[i]-w[i]) @ np.diag(1/np.sqrt(d[i])) for i in range(5)])
lamb, Psi  = np.linalg.eigh(L)[0],np.linalg.eigh(L)[1]
Lamb = np.array([np.diag(lamb[i]) for i in range(5)])    
fhatbar = np.hstack([Psi[i] @ fhat_fiveVTS[:,i] for i in range(5)])
_fhatbar = fhatbar.reshape(5,199)
power = _fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(_fhatbar[i]**2))) for i in range(5)])
fhatbar_threshed = np.where(power_threshed>0,_fhatbar,0)
fhatbarhat = np.array([Psi[i] @ fhatbar_threshed[i] for i in range(5)])    
fhatbarhat_temporal = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_temporal[:,0])
plt.plot(fhatbarhat_temporal[:,1])
plt.plot(fhatbarhat_temporal[:,2])
plt.plot(fhatbarhat_temporal[:,3])
plt.plot(fhatbarhat_temporal[:,4])

2) Spatio

w=np.zeros((5,5))
for i in range(5):
    for j in range(5):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTS.reshape(5,199)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(5)])    
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhatbarhat_spatio = fhatbarhat.reshape(199,-1)
plt.plot(fhatbarhat_spatio[:,0])
plt.plot(fhatbarhat_spatio[:,1])
plt.plot(fhatbarhat_spatio[:,2])
plt.plot(fhatbarhat_spatio[:,3])
plt.plot(fhatbarhat_spatio[:,4])

3) Spatio-Temporal

w=np.zeros((995,995))
for i in range(995):
    for j in range(995):
        if i==j :
            w[i,j] = 0
        elif np.abs(i-j) <= 1 : 
            w[i,j] = 1
d = np.array(w.sum(axis=1))
D = np.diag(d)
L = np.array(np.diag(1/np.sqrt(d)) @ (D-w) @ np.diag(1/np.sqrt(d)))
lamb, Psi = np.linalg.eigh(L)
Lamb = np.diag(lamb)
fhatbar = Psi @ fhat_fiveVTS.reshape(995,1)
power = fhatbar**2 
ebayesthresh = importr('EbayesThresh').ebayesthresh
power_threshed=np.array([np.array(ebayesthresh(FloatVector(fhatbar[i]**2))) for i in range(995)])
fhatbar_threshed = np.where(power_threshed>0,fhatbar,0)
fhatbarhat = Psi @ fhatbar_threshed
fhat_spatio_temporal = fhatbarhat.reshape(199,5,1)
plt.plot(fhat_spatio_temporal[:,0])
plt.plot(fhat_spatio_temporal[:,1])
plt.plot(fhat_spatio_temporal[:,2])
plt.plot(fhat_spatio_temporal[:,3])
plt.plot(fhat_spatio_temporal[:,4])

4. 1로 돌아가서 반복

Original f 진행한 것과 비교

plt.plot((fhatbarhat_temporal[:,2] - fhatbarhat_mean_temporal[:,2])**2)
plt.plot((fhatbarhat_temporal[:,2] - fhatbarhat_random_mean_temporal[:,2])**2)
plt.plot((fhatbarhat_temporal[:,2] - fhatbarhat_twomean_temporal[:,2])**2)

plt.plot((fhatbarhat_temporal[:,2] - fhatbarhat_linearinterpolation_temporal[:,2])**2)
plt.plot((fhatbarhat_temporal[:,2] - fhatbarhat_random_linearinterpolation_temporal[:,2])**2)
plt.plot((fhatbarhat_temporal[:,2] - fhatbarhat_two_linearinterpolation_temporal[:,2])**2)

plt.plot((fhatbarhat_spatio[:,2] - fhatbarhat_mean_spatio[:,2])**2)
plt.plot((fhatbarhat_spatio[:,2] - fhatbarhat_random_mean_spatio[:,2])**2)
plt.plot((fhatbarhat_spatio[:,2] - fhatbarhat_two_mean_spatio[:,2])**2)

plt.plot((fhatbarhat_spatio[:,2] - fhatbarhat_linearinterpolation_spatio[:,2])**2)
plt.plot((fhatbarhat_spatio[:,2] - fhatbarhat_random_linearinterpolation_spatio[:,2])**2)
plt.plot((fhatbarhat_spatio[:,2] - fhatbarhat_two_linearinterpolation_spatio[:,2])**2)

plt.plot((fhat_spatio_temporal[:,2] - fhatbarhat_mean_spatio_temporal[:,2])**2)
plt.plot((fhat_spatio_temporal[:,2] - fhatbarhat_random_mean_spatio_temporal[:,2])**2)
plt.plot((fhat_spatio_temporal[:,2] - fhatbarhat_two_mean_spatio_temporal[:,2])**2)

plt.plot((fhat_spatio_temporal[:,2] - fhat_linearinterpolation_spatio_temporal[:,2])**2)
plt.plot((fhat_spatio_temporal[:,2] - fhat_random_linearinterpolation_spatio_temporal[:,2])**2)
plt.plot((fhat_spatio_temporal[:,2] - fhat_two_linearinterpolation_spatio_temporal[:,2])**2)