resampling data - using SMOTE from imblearn with 3D numpy arrays

Question

I want to resample my dataset. This consists in categorical transformed data with labels of 3 classes. The amount of samples per class are:

• counts of class A: 6945

Answer 1

I am considering a dummy 3d array and assuming a 2d array size by myself,

arr = np.random.rand(160, 10, 25)
orig_shape = arr.shape
print(orig_shape)

Output: (160, 10, 25)

arr = np.reshape(arr, (arr.shape[0], arr.shape[1]))
print(arr.shape)

Output: (4000, 10)

arr = np.reshape(arr, orig_shape))
print(arr.shape)

Output: (160, 10, 25)

Answer 2

I will create each point for a 2-dim array and then reshape it as 3 dim array. I have provided my scripts. If there is any confusion, comment; please reply.

x_train, y_train = zip(*train_dataset)
x_test, y_test = zip(*test_dataset)

dim_1 = np.array(x_train).shape[0]
dim_2 = np.array(x_train).shape[1]
dim_3 = np.array(x_train).shape[2]

new_dim = dim_1 * dim_2

new_x_train = np.array(x_train).reshape(new_dim, dim_3)


new_y_train = []
for i in range(len(y_train)):
    # print(y_train[i])
    new_y_train.extend([y_train[i]]*dim_2)

new_y_train = np.array(new_y_train)

# transform the dataset
oversample = SMOTE()
X_Train, Y_Train = oversample.fit_sample(new_x_train, new_y_train)
# summarize the new class distribution
counter = Counter(Y_Train)
print('The number of samples in TRAIN: ', counter)



x_train_SMOTE = X_Train.reshape(int(X_Train.shape[0]/dim_2), dim_2, dim_3)

y_train_SMOTE = []
for i in range(int(X_Train.shape[0]/dim_2)):
    # print(i)
    value_list = list(Y_Train.reshape(int(X_Train.shape[0]/dim_2), dim_2)[i])
    # print(list(set(value_list)))
    y_train_SMOTE.extend(list(set(value_list)))
    ## Check: if there is any different value in a list 
    if len(set(value_list)) != 1:
        print('\n\n********* STOP: THERE IS SOMETHING WRONG IN TRAIN ******\n\n')
    


dim_1 = np.array(x_test).shape[0]
dim_2 = np.array(x_test).shape[1]
dim_3 = np.array(x_test).shape[2]

new_dim = dim_1 * dim_2

new_x_test = np.array(x_test).reshape(new_dim, dim_3)


new_y_test = []
for i in range(len(y_test)):
    # print(y_train[i])
    new_y_test.extend([y_test[i]]*dim_2)

new_y_test = np.array(new_y_test)

# transform the dataset
oversample = SMOTE()
X_Test, Y_Test = oversample.fit_sample(new_x_test, new_y_test)
# summarize the new class distribution
counter = Counter(Y_Test)
print('The number of samples in TEST: ', counter)



x_test_SMOTE = X_Test.reshape(int(X_Test.shape[0]/dim_2), dim_2, dim_3)

y_test_SMOTE = []
for i in range(int(X_Test.shape[0]/dim_2)):
    # print(i)
    value_list = list(Y_Test.reshape(int(X_Test.shape[0]/dim_2), dim_2)[i])
    # print(list(set(value_list)))
    y_test_SMOTE.extend(list(set(value_list)))
    ## Check: if there is any different value in a list 
    if len(set(value_list)) != 1:
        print('\n\n********* STOP: THERE IS SOMETHING WRONG IN TEST ******\n\n')

Answer 3

from imblearn.over_sampling 
import RandomOverSampler 
import numpy as np 
oversample = RandomOverSampler(sampling_strategy='minority')

X could be a time stepped 3D data like X[sample,time,feature], and y like binary values for each sample. For example: (1,1),(2,1),(3,1) -> 1

X = np.array([[[1,1],[2,1],[3,1]],
             [[2,1],[3,1],[4,1]],
             [[5,1],[6,1],[7,1]],
             [[8,1],[9,1],[10,1]],
             [[11,1],[12,1],[13,1]]
             ])

y = np.array([1,0,1,1,0])

There is no way to train OVERSAMPLER with 3D X values because if you use 2D you will get back 2D data.

Xo,yo = oversample.fit_resample(X[:,:,0], y)
Xo:
[[ 1  2  3]
 [ 2  3  4]
 [ 5  6  7]
 [ 8  9 10]
 [11 12 13]
 [ 2  3  4]]

yo:
[1 0 1 1 0 0]

but if you use 2D data (sample,time,0) to fit the model, it will give back indices, and it is enough to create 3D oversampled data

oversample.fit_resample(X[:,:,0], y)
Xo = X[oversample.sample_indices_]
yo = y[oversample.sample_indices_]

Xo:
[[[ 1  1][ 2  1][ 3  1]]
 [[ 2  1][ 3  1][ 4  1]]
 [[ 5  1][ 6  1][ 7  1]]
 [[ 8  1][ 9  1][10  1]]
 [[11  1][12  1][13  1]]
 [[ 2  1][ 3  1][ 4  1]]]
yo:
[1 0 1 1 0 0]