def mode(data):
""" Substitute missing values with the mode of that column(most frequent).
In the case that there is a tie (there are multiple, most frequent values)
for a column randomly pick one of them.
Parameters
----------
data: numpy.ndarray
Data to impute.
Returns
-------
numpy.ndarray
Imputed data.
"""
nan_xy = matrix.nan_indices(data)
modes = []
for y_i in range(np.shape(data)[1]):
unique_counts = np.unique(data[:, [y_i]], return_counts=True)
max_count = np.max(unique_counts[1])
mode_y = [unique for unique, count in np.transpose(unique_counts)
if count == max_count and not np.isnan(unique)]
modes.append(mode_y) # Appends index of column and column modes
for x_i, y_i in nan_xy:
data[x_i][y_i] = np.random.choice(modes[y_i])
return data
def count_missing(data):
""" Calculate the total percentage of missing values and also the
percentage in each column.
Parameters
----------
data: np.array
Data to impute.
Returns
-------
dict
Percentage of missing values in total and in each column.
"""
size = len(data.flatten())
nan_xy = matrix.nan_indices(data)
np.unique(nan_xy)
counter = {y: 0. for y in np.unique(nan_xy.T[1])}
change_in_percentage = 1. / size
for _, y in nan_xy:
counter[y] += change_in_percentage
total_missing = len(nan_xy) / size
counter["total"] = total_missing
return counter
def describe(data): # verbose=True):
""" Print input/output multiple times
Eventually will be used instead of matrix.nan_indices everywhere
Parameters
----------
data: numpy.nd.array
The data you want to get a description from
verbose: boolean(optional)
Decides whether the description is short or long form
Returns
-------
dict
missingness: list
Confidence interval of data being MCAR, MAR or MNAR - in that order
nan_xy: list of tuples
Indices of all null points
nan_n: list
Total number of null values for each column
pmissing_n: float
Percentage of missing values in dataset
nan_rows: list
Indices of all rows that are completely null
nan_cols: list
Indices of all columns that are completely null
mean_rows: list
Mean value of each row
mean_cols: list
Mean value of each column
std_dev: list
std dev for each row/column
min_max: list
Finds the minimum and maximum for each row
"""
# missingness = [0.33, 0.33, 0.33] # find_missingness(data)
nan_xy = matrix.nan_indices(data)
nan_n = len(nan_xy)
pmissing_n = float(nan_n / len(data.flatten))
# pmissing_rows = ""
# pmissing_cols = ""
# nan_rows = ""
# nan_cols = ""
# mean_rows = ""
# mean_cols = ""
# std_dev = ""
# "missingness": missingness,
description = {"nan_xy": nan_xy, "nan_n": nan_n, "pmissing_n": pmissing_n}
# "pmissing_rows": pmissing_rows,
# "pmissing_cols": pmissing_cols,
# "nan_rows": nan_rows,
# "nan_cols": nan_cols,
# "mean_rows": mean_rows,
# "mean_cols": mean_cols,
# "std_dev": std_dev}
return description
def em(data, loops=50):
""" Imputes given data using expectation maximization.
E-step: Calculates the expected complete data log likelihood ratio.
M-step: Finds the parameters that maximize the log likelihood of the
complete data.
Parameters
----------
data: numpy.nd.array
Data to impute.
loops: int
Number of em iterations to run before breaking.
inplace: boolean
If True, operate on the numpy array reference
Returns
-------
numpy.nd.array
Imputed data.
"""
nan_xy = matrix.nan_indices(data)
for x_i, y_i in nan_xy:
col = data[:, int(y_i)]
mu = col[~np.isnan(col)].mean()
std = col[~np.isnan(col)].std()
col[x_i] = np.random.normal(loc=mu, scale=std)
previous, i = 1, 1
for i in range(loops):
# Expectation
mu = col[~np.isnan(col)].mean()
std = col[~np.isnan(col)].std()
# Maximization
col[x_i] = np.random.normal(loc=mu, scale=std)
# Break out of loop if likelihood doesn't change at least 10%
# and has run at least 5 times
delta = (col[x_i] - previous) / previous
if i > 5 and delta < 0.1:
data[x_i][y_i] = col[x_i]
break
data[x_i][y_i] = col[x_i]
previous = col[x_i]
return data
def mean(data):
""" Substitute missing values with the mean of that column.
Parameters
----------
data: numpy.ndarray
Data to impute.
Returns
-------
numpy.ndarray
Imputed data.
"""
nan_xy = matrix.nan_indices(data)
for x_i, y_i in nan_xy:
row_wo_nan = data[:, [y_i]][~np.isnan(data[:, [y_i]])]
new_value = np.mean(row_wo_nan)
data[x_i][y_i] = new_value
return data
def random(data):
""" Fill missing values in with a randomly selected value from the same
column.
Parameters
----------
data: numpy.ndarray
Data to impute.
Returns
-------
numpy.ndarray
Imputed data.
"""
nan_xy = matrix.nan_indices(data)
for x, y in nan_xy:
uniques = np.unique(data[:, y])
uniques = uniques[~np.isnan(uniques)]
data[x][y] = np.random.choice(uniques)
return data
def median(data):
""" Substitute missing values with the median of that column(middle).
Parameters
----------
data: numpy.ndarray
Data to impute.
Returns
-------
numpy.ndarray
Imputed data.
"""
nan_xy = matrix.nan_indices(data)
cols_missing = set(nan_xy.T[1])
medians = {}
for y_i in cols_missing:
cols_wo_nan = data[:, [y_i]][~np.isnan(data[:, [y_i]])]
median_y = np.median(cols_wo_nan)
medians[str(y_i)] = median_y
for x_i, y_i in nan_xy:
data[x_i][y_i] = medians[str(y_i)]
return data
def test_missing_values_present():
""" Check that the dataset is corrupted (missing values present)"""
assert matrix.nan_indices(data).size != 0
def moving_window(data,
nindex=None,
wsize=5,
errors="coerce",
func=np.mean,
inplace=False):
""" Interpolate the missing values based on nearby values.
For example, with an array like this:
array([[-1.24940, -1.38673, -0.03214945, 0.08255145, -0.007415],
[ 2.14662, 0.32758 , -0.82601414, 1.78124027, 0.873998],
[-0.41400, -0.977629, nan, -1.39255344, 1.680435],
[ 0.40975, 1.067599, 0.29152388, -1.70160145, -0.565226],
[-0.54592, -1.126187, 2.04004377, 0.16664863, -0.010677]])
Using a `k` or window size of 3. The one missing value would be set
to -1.18509122. The window operates on the horizontal axis.
Usage
-----
The parameters default the function to a moving mean. You may want to change
the default window size:
moving_window(data, wsize=10)
To only look at past data (null value is at the rightmost index in the window):
moving_window(data, nindex=-1)
To use a custom function:
moving_window(data, func=np.median)
You can also do something like take 1.5x the max of previous values in the window:
moving_window(data, func=lambda arr: max(arr) * 1.50, nindex=-1)
Parameters
----------
data: numpy.ndarray
2D matrix to impute.
nindex: int
Null index. Index of the null value inside the moving average window.
Use cases: Say you wanted to make value skewed toward the left or right
side. 0 would only take the average of values from the right and -1
would only take the average of values from the left
wsize: int
Window size. Size of the moving average window/area of values being used
for each local imputation. This number includes the missing value.
errors: {"raise", "coerce", "ignore"}
Errors will occur with the indexing of the windows - for example if there
is a nan at data[x][0] and `nindex` is set to -1 or there is a nan at
data[x][-1] and `nindex` is set to 0. `"raise"` will raise an error,
`"coerce"` will try again using an nindex set to the middle and `"ignore"`
will just leave it as a nan.
inplace: {True, False}
Whether to return a copy or run on the passed-in array
Returns
-------
numpy.ndarray
Imputed data.
"""
if errors == "ignore":
raise Exception("`errors` value `ignore` not implemented yet. Sorry!")
if not inplace:
data = data.copy()
if nindex is None: # If using equal window side lengths
assert wsize % 2 == 1, "The parameter `wsize` should not be even "\
"if the value `nindex` is not set since it defaults to the midpoint "\
"and an even `wsize` makes the midpoint ambiguous"
wside_left = wsize // 2
wside_right = wsize // 2
else: # If using custom window side lengths
assert nindex < wsize, "The null index must be smaller than the window size"
if nindex == -1:
wside_left = wsize - 1
wside_right = 0
else:
wside_left = nindex
wside_right = wsize - nindex - 1
while True:
nan_xy = matrix.nan_indices(data)
n_nan_prev = len(nan_xy)
for x_i, y_i in nan_xy:
left_i = max(0, y_i - wside_left)
right_i = min(len(data), y_i + wside_right + 1)
window = data[x_i, left_i:right_i]
window_not_null = window[~np.isnan(window)]
if len(window_not_null) > 0:
try:
data[x_i][y_i] = func(window_not_null)
continue
except Exception as e:
if errors == "raise":
raise e
if errors == "coerce":
# If either the window has a length of 0 or the aggregate function fails somehow,
# do a fallback of just trying the best we can by using it as the middle and trying
# to recalculate. Use temporary wside_left/wside_right, for only the calculation of
# this specific problamatic value
wside_left_tmp = wsize // 2
wside_right_tmp = wside_left_tmp
left_i_tmp = max(0, y_i - wside_left_tmp)
right_i_tmp = min(len(data), y_i + wside_right_tmp + 1)
window = data[x_i, left_i_tmp:right_i_tmp]
window_not_null = window[~np.isnan(window)]
try:
data[x_i][y_i] = func(window_not_null)
except Exception as e:
print("Exception:", e)
if n_nan_prev == len(matrix.nan_indices(data)):
break
return data
def fast_knn(data,
k=3,
eps=0,
p=2,
distance_upper_bound=np.inf,
leafsize=10,
idw_fn=idw.shepards,
init_impute_fn=mean):
""" Impute using a variant of the nearest neighbours approach
Basic idea: Impute array with a passed in initial impute fn (mean impute)
and then use the resulting complete array to construct a KDTree. Use this
KDTree to compute nearest neighbours. After finding `k` nearest
neighbours, take the weighted average of them. Basically, find the nearest
row in terms of distance
This approach is much, much faster than the other implementation (fit+transform
for each subset) which is almost prohibitively expensive.
Parameters
----------
data: numpy.ndarray
2D matrix to impute.
k: int, optional
Parameter used for method querying the KDTree class object. Number of
neighbours used in the KNN query. Refer to the docs for
[`scipy.spatial.KDTree.query`]
(https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.KDTree.query.html).
eps: nonnegative float, optional
Parameter used for method querying the KDTree class object. From the
SciPy docs: "Return approximate nearest neighbors; the kth returned
value is guaranteed to be no further than (1+eps) times the distance to
the real kth nearest neighbor". Refer to the docs for
[`scipy.spatial.KDTree.query`]
(https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.KDTree.query.html).
p : float, 1<=p<=infinity, optional
Parameter used for method querying the KDTree class object. Straight from the
SciPy docs: "Which Minkowski p-norm to use. 1 is the
sum-of-absolute-values Manhattan distance 2 is the usual Euclidean
distance infinity is the maximum-coordinate-difference distance". Refer to
the docs for
[`scipy.spatial.KDTree.query`]
(https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.KDTree.query.html).
distance_upper_bound : nonnegative float, optional
Parameter used for method querying the KDTree class object. Straight
from the SciPy docs: "Return only neighbors within this distance. This
is used to prune tree searches, so if you are doing a series of
nearest-neighbor queries, it may help to supply the distance to the
nearest neighbor of the most recent point." Refer to the docs for
[`scipy.spatial.KDTree.query`]
(https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.KDTree.query.html).
leafsize: int, optional
Parameter used for construction of the `KDTree` class object. Straight from
the SciPy docs: "The number of points at which the algorithm switches
over to brute-force. Has to be positive". Refer to the docs for
[`scipy.spatial.KDTree`](https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.spatial.KDTree.html)
for more information.
idw_fn: fn, optional
Function that takes one argument, a list of distances, and returns weighted percentages. You can define a custom
one or bootstrap from functions defined in `impy.util.inverse_distance_weighting` which can be using
functools.partial, for example: `functools.partial(impy.util.inverse_distance_weighting.shepards, power=1)`
init_impute_fn: fn, optional
Returns
-------
numpy.ndarray
Imputed data.
Examples
--------
>>> data = np.arange(25).reshape((5, 5)).astype(np.float)
>>> data[0][2] = np.nan
>>> data
array([[ 0., 1., nan, 3., 4.],
[ 5., 6., 7., 8., 9.],
[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.],
[20., 21., 22., 23., 24.]])
>> fast_knn(data, k=1) # Weighted average (by distance) of nearest 1 neighbour
array([[ 0., 1., 7., 3., 4.],
[ 5., 6., 7., 8., 9.],
[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.],
[20., 21., 22., 23., 24.]])
>> fast_knn(data, k=2) # Weighted average of nearest 2 neighbours
array([[ 0. , 1. , 10.08608891, 3. , 4. ],
[ 5. , 6. , 7. , 8. , 9. ],
[10. , 11. , 12. , 13. , 14. ],
[15. , 16. , 17. , 18. , 19. ],
[20. , 21. , 22. , 23. , 24. ]])
>> fast_knn(data, k=3)
array([[ 0. , 1. , 13.40249283, 3. , 4. ],
[ 5. , 6. , 7. , 8. , 9. ],
[10. , 11. , 12. , 13. , 14. ],
[15. , 16. , 17. , 18. , 19. ],
[20. , 21. , 22. , 23. , 24. ]])
>> fast_knn(data, k=5) # There are at most only 4 neighbours. Raises error
...
IndexError: index 5 is out of bounds for axis 0 with size 5
"""
nan_xy = matrix.nan_indices(data)
data_c = init_impute_fn(data)
kdtree = KDTree(data_c, leafsize=leafsize)
for x_i, y_i in nan_xy:
distances, indices = kdtree.query(
data_c[x_i],
k=k + 1,
eps=eps,
p=p,
distance_upper_bound=distance_upper_bound)
# Will always return itself in the first index. Delete it.
distances, indices = distances[1:], indices[1:]
# Add small constant to distances to avoid division by 0
distances += 1e-3
weights = idw_fn(distances)
# Assign missing value the weighted average of `k` nearest neighbours
data[x_i][y_i] = np.dot(weights, [data_c[ind][y_i] for ind in indices])
return data
def buck_iterative(data):
""" Iterative variant of buck's method
- Variable to regress on is chosen at random.
- EM type infinite regression loop stops after change in prediction from
previous prediction < 10% for all columns with missing values
A Method of Estimation of Missing Values in Multivariate Data Suitable for
use with an Electronic Computer S. F. Buck Journal of the Royal Statistical
Society. Series B (Methodological) Vol. 22, No. 2 (1960), pp. 302-306
Parameters
----------
data: numpy.ndarray
Data to impute.
Returns
-------
numpy.ndarray
Imputed data.
"""
nan_xy = matrix.nan_indices(data)
# Add a column of zeros to the index values
nan_xyz = np.append(nan_xy, np.zeros((np.shape(nan_xy)[0], 1)), axis=1)
nan_xyz = [[int(x), int(y), v] for x, y, v in nan_xyz]
temp = []
cols_missing = {y for _, y, _ in nan_xyz}
# Step 1: Simple Imputation, these are just placeholders
for x_i, y_i, value in nan_xyz:
# Column containing nan value without the nan value
col = data[:, [y_i]][~np.isnan(data[:, [y_i]])]
new_value = np.mean(col)
data[x_i][y_i] = new_value
temp.append([x_i, y_i, new_value])
nan_xyz = temp
# Step 5: Repeat step 2 - 4 until convergence (the 100 is arbitrary)
converged = [False] * len(nan_xyz)
while not all(converged):
# Step 2: Placeholders are set back to missing for one variable/column
dependent_col = int(np.random.choice(list(cols_missing)))
missing_xs = [int(x) for x, y, value in nan_xyz if y == dependent_col]
# Step 3: Perform linear regression using the other variables
x_train, y_train = [], []
for x_i in (x_i for x_i in range(len(data)) if x_i not in missing_xs):
x_train.append(np.delete(data[x_i], dependent_col))
y_train.append(data[x_i][dependent_col])
model = LinearRegression()
model.fit(x_train, y_train)
# Step 4: Missing values for the missing variable/column are replaced
# with predictions from our new linear regression model
# For null indices with the dependent column that was randomly chosen
for i, z in enumerate(nan_xyz):
x_i = z[0]
y_i = z[1]
value = data[x_i, y_i]
if y_i == dependent_col:
# Row 'x' without the nan value
new_value = model.predict(
[np.delete(data[x_i], dependent_col)])
data[x_i][y_i] = new_value.reshape(1, -1)
if value == 0.0:
delta = (new_value - value) / 0.01
else:
delta = (new_value - value) / value
converged[i] = abs(delta) < 0.1
return data
