Filling missing time values in a multi-indexed dataframe

Question

Problem and what I want

I have a data file that comprises time series read asynchronously from multiple sensors. Basically for every data element in my file, I have a sensor ID and time at which it was read, but I do not always have all sensors for every time, and read times may not be evenly spaced. Something like:

ID,time,data
0,0,1
1,0,2
2,0,3
0,1,4
2,1,5  # skip some sensors for some time steps
0,2,6
2,2,7
2,3,8
1,5,9  # skip some time steps
2,5,10

Important note the actual time column is of datetime type.

What I want is to be able to zero-order hold (forward fill) values for every sensor for any time steps where that sensor does not exist, and either set to zero or back fill any sensors that are not read at the earliest time steps. What I want is a dataframe that looks like it was read from:

ID,time,data
0,0,1
1,0,2
2,0,3
0,1,4
1,1,2  # ID 1 hold value from time step 0
2,1,5
0,2,6
1,2,2  # ID 1 still holding
2,2,7
0,3,6  # ID 0 holding
1,3,2  # ID 1 still holding
2,3,8
0,5,6  # ID 0 still holding, can skip totally missing time steps
1,5,9  # ID 1 finally updates
2,5,10

Pandas attempts so far

I initialize my dataframe and set my indices:

df = pd.read_csv(filename, dtype=np.int)
df.set_index(['ID', 'time'], inplace=True)

I try to mess with things like:

filled = df.reindex(method='ffill')

or the like with various values passed to the index keyword argument like df.index, ['time'], etc. This always either throws an error because I passed an invalid keyword argument, or does nothing visible to the dataframe. I think it is not recognizing that the data I am looking for is "missing".

I also tried:

df.update(df.groupby(level=0).ffill())

or level=1 based on Multi-Indexed fillna in Pandas, but I get no visible change to the dataframe again, I think because I don't have anything currently where I want my values to go.

Numpy attempt so far

I have had some luck with numpy and non-integer indexing using something like:

data = [np.array(df.loc[level].data) for level in df.index.levels[0]]
shapes = [arr.shape for arr in data]
print(shapes)
# [(3,), (2,), (5,)]
data = [np.array([arr[i] for i in np.linspace(0, arr.shape[0]-1, num=max(shapes)[0])]) for arr in data]
print([arr.shape for arr in data])
# [(5,), (5,), (5,)]

But this has two problems:

1. It takes me out of the pandas world, and I now have to manually maintain my sensor IDs, time index, etc. along with my feature vector (the actual data column is not just one column but a ton of values from a sensor suite).
2. Given the number of columns and the size of the actual dataset, this is going to be clunky and inelegant to implement on my real example. I would prefer a way of doing it in pandas.

The application

Ultimately this is just the data-cleaning step for training recurrent neural network, where for each time step I will need to feed a feature vector that always has the same structure (one set of measurements for each sensor ID for each time step).

Thank you for your help!

Answer 1

Here is one way , by using reindex and category

df.time=df.time.astype('category',categories =[0,1,2,3,4,5])
new_df=df.groupby('time',as_index=False).apply(lambda x : x.set_index('ID').reindex([0,1,2])).reset_index()
new_df['data']=new_df.groupby('ID')['data'].ffill()
new_df.drop('time',1).rename(columns={'level_0':'time'})
Out[311]: 
    time  ID  data
0      0   0   1.0
1      0   1   2.0
2      0   2   3.0
3      1   0   4.0
4      1   1   2.0
5      1   2   5.0
6      2   0   6.0
7      2   1   2.0
8      2   2   7.0
9      3   0   6.0
10     3   1   2.0
11     3   2   8.0
12     4   0   6.0
13     4   1   2.0
14     4   2   8.0
15     5   0   6.0
16     5   1   9.0
17     5   2  10.0

Answer 2

You can have a dictionary of last readings for each sensors. You'll have to pick some initial value; the most logical choice is probably to back-fill the earliest reading to earlier times. Once you've populated your last_reading dictionary, you can just sort all the readings by time, update the dictionary for each reading, and then fill in rows according to the dictionay. So after you have your last_reading dictionary initialized:

last_time = readings[1][time]
for reading in readings:
   if reading[time] > last_time:
      for ID in ID_list:
         df.loc[last_time,ID] = last_reading[ID]
      last_time = reading[time]
   last_reading[reading[ID]] = reading[data]
#the above for loop doesn't update for the last time
#so you'll have to handle that separately
for ID in ID_list:
    df.loc[last_time,ID] = last_reading[ID]
    last_time = reading[time]

This assumes that you have only one reading for each time/sensor pair, and that 'readings' a list of dictionaries sorted by time. It also assumes that df has the different sensors as columns and different times as index. Adjust the code as necessary if otherwise. You can also probably optimize it a bit more by updating a whole row at once instead of using a for loop, but I didn't want to deal with making sure I had the Pandas syntax right.

Looking at the application, though, you might want to have each cell in the dataframe be not a number but a tuple of last value and time it was read, so replace last_reading[reading[ID]] = reading[data] with last_reading[reading[ID]] = [reading[data],reading[time]]. Your neural net can then decide how to weight data based on how old it is.

Answer 3

I got this to work with the following, which I think is pretty general for any case like this where the time index for which you want to fill values is the second in a multi-index with two indices:

# Remove duplicate time indices (happens some in the dataset, pandas freaks out).
df = df[~df.index.duplicated(keep='first')]

# Unstack the dataframe and fill values per serial number forward, backward.
df = df.unstack(level=0)
df.update(df.ffill())  # first ZOH forward
df.update(df.bfill())  # now back fill values that are not seen at the beginning

# Restack the dataframe and re-order the indices.
df = df.stack(level=1)
df = df.swaplevel()

This gets me what I want, although I would love to be able to keep the duplicate time entries if anybody knows of a good way to do this.

You could also use df.update(df.fillna(0)) instead of backfilling if starting unseen values at zero is preferable for a particular application.

I put the above code block in a function called clean_df that takes the dataframe as argument and returns the cleaned dataframe.