Data partitioning based on timeseries and groups in R-CodePudding

I need to split a large dataframe of meterological timeseries into a training and validation samples. It contains data from multiple stations, which have varying period of observations. How could I divide it so that proportion of training and validation observations is equal across each station. Given the following dataset:

Station	Date	temp
A	2012-01-01	-0.8
A	2012-01-02	0.1
A	2012-01-03	0.5
A	2012-01-04	0.4
B	2012-01-01	0.1
B	2012-01-02	0.5

and assuming that the training set should include only first 50% of the observations from each station, the desired output would be:

Station	Date	temp
A	2012-01-01	-0.8
A	2012-01-02	0.1
B	2012-01-01	0.1

CodePudding user response：

Given your example you could use slice_head from dplyr. For creating the validation, remove the records that are in training. This to avoid selecting duplictates in case there is an uneven number of records for a station.

training <- df1 %>% 
  mutate(Date = as.Date(Date),
         id = row_number()) %>% 
  group_by(Station) %>% 
  slice_head(prop = 0.5)
  
validation <- df1 %>% 
  mutate(Date = as.Date(Date),
         id = row_number()) %>%
  filter(!id %in% training$id)

training
# A tibble: 3 x 4
# Groups:   Station [2]
  Station Date        temp    id
  <chr>   <date>     <dbl> <int>
1 A       2012-01-01  -0.8     1
2 A       2012-01-02   0.1     2
3 B       2012-01-01   0.1     5

validation
  Station       Date temp id
1       A 2012-01-03  0.5  3
2       A 2012-01-04  0.4  4
3       B 2012-01-02  0.5  6