I have a data frame with user IDs all in one column and "time series" for each user, which looks like this:
df = pd.DataFrame({'user_id': [1, 1, 1, 2, 2, 2, 3, 3, 3], 'time': [0, 1, 3, 4, 8, 10, 20, 30, 80], 'score': [1, 2, 3, 4, 5, 6, 7, 8, 9]})
I want to calculate time differences for each user_id:
df = pd.DataFrame({'user_id': [1, 1, 1, 2, 2, 2, 3, 3, 3], 'time': [0, 1, 2, 4, 4, 2, 20, 10, 50], 'score': [1, 2, 3, 4, 5, 6, 7, 8, 9]})
I think np.diff would work if I could limit it to each user_id. This is my first question on StackOverflow, hope my question is clear enough.
Try, using groupby with diff, then fillna first NaN with current time value:
df['diff'] = df.groupby('user_id')['time'].diff().fillna(df['time'])
Output:
user_id time score diff
0 1 0 1 0.0
1 1 1 2 1.0
2 1 3 3 2.0
3 2 4 4 4.0
4 2 8 5 4.0
5 2 10 6 2.0
6 3 20 7 20.0
7 3 30 8 10.0
8 3 80 9 50.0
Related
For example I have DataFrame
df = pd.DataFrame({'a': [1, 2, 3, 4, 5, 6, 7, 8, 9], 'b': [2, 2, 4, 3, 1000, 2000, 1, 500, 3]})
I need to cut by outliers and get these intervals: 1-4, 5-6, 7, 8, 9.
Cutting with pd.cut and pd.qcut does not give these results
You can group them by consecutive values depending on the above/below mask:
m = df['b'].gt(100)
df['group'] = m.ne(m.shift()).cumsum()
output:
a b group
0 1 2 1
1 2 2 1
2 3 4 1
3 4 3 1
4 5 1000 2
5 6 2000 2
6 7 1 3
7 8 500 4
8 9 3 5
I am trying to calculate the rolling average of 4 days on the below dataset. The result should also be calculated based on group of 2 other columns.
For example:
df_time = pd.DataFrame({'A': [123, 123, 278, 278, 278, 123, 345, 278, 123,278, 278],
'B': [1, 2, 2, 2, 2, 1, 3, 2, 1, 2, 2],
'C': [0, 4, 2, 0, 4, 5, 3, 2, 1, 3, 2],
'D' : [pd.Timestamp('20130101'),
pd.Timestamp('20130102'),
pd.Timestamp('20130101'),
pd.Timestamp('20130102'),
pd.Timestamp('20130103'),
pd.Timestamp('20130103'),
pd.Timestamp('20130104'),
pd.Timestamp('20130104'),
pd.Timestamp('20130105'),
pd.Timestamp('20130106'),
pd.Timestamp('20130109')],
'rol_avg': [0, 2, 2, 1, 2, 5, 3, 3, 3, 2, 2.5] } )
df_time.sort_values(by='D')
sum_df = df_time.groupby(by=['A', 'B', 'D'])['C'].sum()
rolling_average_series = sum_df.groupby(level=[0, 1]).rolling(4, min_periods=1).mean()
returns
It is finding the rolling average of the past 4 occurrences of columns A and B. But what I need is rolling average of 4 days with respect to the Date column 'D' and grouped by column 'A' and 'B'
Here, for A=278 and B=2 for 09-01-2013 we have only 06-01-2013 which falls in 4days rolling average so the average should be (2+3)/2 = 2.5
How do I implement this functionality?
Any help is greatly appreciated
I hope I've understood your question well. You can specify column with on= in .rolling() and then specify window="4D" - that you want 4 days rolling window:
df_time = df_time.sort_values(by="D")
sum_df = df_time.groupby(by=["A", "B", "D"], as_index=False)["C"].sum()
rolling_average_series = (
sum_df.groupby(["A", "B"])
.rolling(on="D", window="4D", min_periods=1)
.mean()
)
print(rolling_average_series)
Prints:
D C
A B
123 1 0 2013-01-01 0.0
1 2013-01-03 2.5
2 2013-01-05 3.0
2 3 2013-01-02 4.0
278 2 4 2013-01-01 2.0
5 2013-01-02 1.0
6 2013-01-03 2.0
7 2013-01-04 2.0
8 2013-01-06 3.0
9 2013-01-09 2.5
345 3 10 2013-01-04 3.0
I'm looking to feed an autoencoder my features, as both training and target data. Majority of the features have rank 1; are a single column of values like [1,2,3,4]. Some have been put through one hot encoding so the tensors are of rank 2 and have X columns, with X being the number of categorical values in the one hot encoder, so something like:
['a', 'a, 'b', c'] -> [[1,0,0], [1,0,0], [0,1,0], [0,0,1]]
For some reason keras's Model.fit don't accept y values if the training data are generators or datasets. So I have to provide my training data as a tuple of (features, targets), and in this case targets=features, but at the same time, features is a dictionary of tensors so I must concatenate all of the tensors in features into a single tensor.
I can do tf.concat across all of my feature columns except the one-hot encoded columns, which have rank 2 (instead of 1). How can I somehow turn the one-hot encoded features in X individual tensors and then concat them together?
Same issue here where the OP solved his issue using tf.concat, but I can't do that here.
I was trying to comment your question to ask for a couple of details, but I can't do that because I don't have reputation score (I am new to stack overflow and this is my first post ever). So I will try to explain based on what I understood of your problem - hope it helps.
You can use Pandas to hot-encode your categorical features.
Example Dataset
import pandas as pd
import random
import numpy as np
dataset = {'feature_A': np.random.randint(10, size=10),
'feature_B': np.random.randint(10, size=10),
'feature_C': np.random.randint(10, size=10),
'categorical_feature': np.array([chr(97 + random.randint(0, 2)) for i in range(10)])}
print(dataset)
{'feature_A': array([6, 8, 2, 0, 4, 8, 6, 4, 3, 8]),
'feature_B': array([0, 6, 8, 6, 7, 3, 4, 6, 1, 6]),
'feature_C': array([0, 7, 7, 3, 7, 4, 0, 3, 2, 7]),
'categorical_feature': array(['a', 'a', 'c', 'b', 'c', 'c', 'a', 'a', 'c', 'c'], dtype='<U1')}
Pandas DataFrame
Transform the dataset into a Pandas DataFrame
df = pd.DataFrame(dataset)
print(df)
Feature_A Feature_B Feature_C Categorical_Feature
0 1 9 9 e
1 9 9 8 c
2 5 8 3 c
3 3 5 7 c
4 9 8 10 d
5 6 9 6 c
6 1 4 5 d
7 9 5 2 e
8 3 9 2 c
9 3 8 10 a
One-hot & concatenate
One-hot encode the categorical features and concatenate to the main DataFrame (and drop original categorical feature column)
df = pd.concat((df, pd.get_dummies(df['categorical_feature'])), axis=1).drop('categorical_feature', axis=1)
print(df)
feature_A feature_B feature_C a b c
0 0 0 1 0 0 1
1 5 7 0 0 0 1
2 9 2 6 0 1 0
3 1 1 4 0 0 1
4 5 8 8 0 0 1
5 9 1 8 1 0 0
6 5 6 8 1 0 0
7 9 5 0 1 0 0
8 9 4 5 1 0 0
9 8 3 5 1 0 0
NumPy array
Then you can simply get the values of the DataFrame as a numpy array by using the attribute .values. Each row now is one training example that comprises of all the value features + the categorical features as hot encoded vector.
You can use the numpy array directly into your model or, if you wish, you can also transform it into tensorflow tensor by using tf.data.Dataset.from_tensor_slices().
dataset = df.values
print(dataset)
array([[0, 0, 1, 0, 0, 1],
[5, 7, 0, 0, 0, 1],
[9, 2, 6, 0, 1, 0],
[1, 1, 4, 0, 0, 1],
[5, 8, 8, 0, 0, 1],
[9, 1, 8, 1, 0, 0],
[5, 6, 8, 1, 0, 0],
[9, 5, 0, 1, 0, 0],
[9, 4, 5, 1, 0, 0],
[8, 3, 5, 1, 0, 0]], dtype=int32)
I have a data frame like this
df_a = pd.DataFrame({'a': [2, 4, 5, 6, 12],
'b': [3, 5, 7, 9, 15]})
Out[112]:
a b
0 2 3
1 4 5
2 5 7
3 6 9
4 12 15
and mean out
df_a.mean()
Out[118]:
a 5.800
b 7.800
dtype: float64
I want this;
df_a[df_a.index.isin([3, 4])] = df.mean()
But I'm getting an error. How do I achieve this?
I gave an example here. There are observations that I need to change a lot in the data that I am working with. And I keep their index values in a list
If you want to overwrite the values of rows in a list, you can do it with iloc
df_a = pd.DataFrame({'a': [2, 4, 5, 6, 12], 'b': [3, 5, 7, 9, 15]})
idx_list = [3, 4]
df_a.iloc[idx_list,:] = df_a.mean()
Output
a b
0 2.0 3.0
1 4.0 5.0
2 5.0 7.0
3 5.8 7.8
4 5.8 7.8
edit
If you're using an older version of pandas and see NaNs instead of wanted values, you can use a for loop
df_a_mean = df_a.mean()
for i in idx_list:
df_a.iloc[i,:] = df_a_mean
Is it possible to change a single level column dataframe to a multi-column dataframe? If we have a dataframe like this,
import pandas as pd
df = pd.DataFrame({
'a': [0, 1, 2, 3],
'b': [4, 5, 6, 7],
'c': [3, 5, 6, 2],
'd': [1, 5, 7, 0],
})
can we change it's column names as below?. So, briefly what I am trying to do is to have 2-levels of column index without changing the values of the dataframe.
A B
a b c d
0 0 4 3 1
1 1 5 5 5
2 2 6 6 7
3 3 7 2 0
Any help?
IIUC, use pd.MultiIndex.from_tuples to create multiindex header and assign to the dataframe.columns:
df = pd.DataFrame({
'a': [0, 1, 2, 3],
'b': [4, 5, 6, 7],
'a2': [3, 5, 6, 2],
'b2': [1, 5, 7, 0],
})
df.columns=pd.MultiIndex.from_tuples([('A','a'),('A','b'),('B','c'),('B','d')])
df
Output:
A B
a b c d
0 0 4 3 1
1 1 5 5 5
2 2 6 6 7
3 3 7 2 0