Issues with nested table using purrr package--converted TS objects have datatype character instead of dbl

[princessblackswan]

Hi all,

Hoping someone with experience with nested table to help me out.

I have nested table with 45 rows, and I would like to convert train data to time series objects ts and msts for fitting. So that has been done with no issue, however I notice that the params list contains character type S3:ts objects, which is out of the ordinary because the previous times I have done this the datatype is usually dbl. As a result I'm unable to fit the models, because of this non-numeric TS objects. I tried a workaround to use lapply for all elements inside the list to convert them to numeric, but when this is done its no longer a time series object.

Hoping someone can help! Thank you

#Nest the new tables
walmart_improve_nested <- walmart_data_improve %>%
  group_by(Store) %>%
  nest()

#Cross Validation
walmart_improve_nested <- walmart_improve_nested %>%
  mutate(train = map(data, train_split),
         test = map(data, test_split))

#Decompose and plot
walmart_ts_decompose <- function(df){
  df$Weekly_Sales = as.numeric(df$Weekly_Sales)
  ts(df$Weekly_Sales, frequency=52) %>% decompose() %>% autoplot()
}

walmart_msts_decompose <- function(df){
  df$Weekly_Sales = as.numeric(df$Weekly_Sales)
  msts(df$Weekly_Sales, seasonal.periods = c(4, 12, 26, 52)) %>% mstl() %>% autoplot()
}

#Visualize decomposed plots 
walmart_improve_nested <- walmart_improve_nested %>% 
  mutate(ts_decompose = map(train, walmart_ts_decompose),
         msts_decompose = map(train, walmart_msts_decompose))

#Create list of time series object functions
ts_function <- list(
  ts_obj = function(x) ts(as.numeric(x$Weekly_Sales), frequency=52),
  msts_obj = function(x) msts(as.numeric(x$Weekly_Sales), seasonal.periods = c(4,12,26,52))
)


#As there are two objects, replicate each unique Store number twice i.e one for each time series object
ts_function<- ts_function %>% 
  rep(length(unique(walmart$Store))) %>% 
  enframe("ts_func_name", "ts_func") %>% 
  mutate(Store = sort(rep(unique(walmart$Store), length(unique(.$ts_func_name)))))


#left join 
walmart_improve_nested <- left_join(walmart_improve_nested, ts_function) 

model_function <- list(
   stlm = function(x) stlm(x),
   tbats = function(x) tbats(x),
   HoltWinters = function(x) HoltWinters(x)
   )

model_function<- model_function %>% 
  rep(length(unique(walmart$Store))) %>% 
  enframe("model_name", "model_func") %>% 
  mutate(Store = sort(rep(unique(walmart$Store), length(unique(.$model_name)))))

#left join 
walmart_improve_nested <- left_join(walmart_improve_nested, model_function) 

walmart_improve_nested <- walmart_improve_nested %>%
  mutate(params = map(train, ~list(x = .x)))

walmart_improve_nested <- walmart_improve_nested %>% 
  mutate(
    data = invoke_map(ts_func, params),
    params = map(data, ~list(x = .x))
  )

The above code runs fine and are as expected with the exception of the class of the params list, which is character instead of dbl like usual. I have explicitly coerce the target variable as numeric but somehow the resulting TS object becomes character. I believe this is the cause of the problem as the other times when the type of params is 'dbl' there has been no issues.

Some screenshots

Please note that the following code results in error:

walmart_improve_nested <- walmart_improve_nested %>% mutate( fitted = invoke_map(model_func, params) )
Error message:

Error in mutate():
ℹ In argument: fitted = invoke_map(model_func, params).
ℹ In group 1: Store = "1".
Caused by error in map2():
ℹ In index: 1.
Caused by error in +.default:
! non-numeric argument to binary operator
Backtrace:

1. walmart_improve_nested %>% ...
2. global <fn>(x = <ts>)
3. forecast::stlm(x)
4. forecast::mstl(x, s.window = s.window, robust = robust)
5. base::NextMethod(.Generic)

Thank you

[IrfanChairurrachman]

Dear @princessblackswan ,

Pertanyaan Anda sudah kami baca dan sedang kami telusuri permasalahannya. Harap menunggu untuk balasan selanjutnya, terima kasih banyak.

[hwulanayu]

It seems like the issue might be arising from the structure of the params list. Specifically, when you apply the invoke_map function to model_func and params, it might not be handling the character type of your time series objects correctly.

To resolve this, you can modify the code to ensure that the params list contains numeric time series objects. You can do this by directly applying your time series functions (ts_obj and msts_obj) to the data within the train list. Here's a modified version of your code:

walmart_improve_nested <- walmart_improve_nested %>%
  mutate(params = map(train, ts_function$ts_obj))

# Continue with the rest of your code

walmart_improve_nested <- walmart_improve_nested %>%
  mutate(
    data = invoke_map(ts_func, params),
    params = map(data, ~list(x = .x))
  )
...

This modification directly applies ts_function$ts_obj to the train list, ensuring that the resulting params list contains numeric time series objects. After making this change, you should be able to proceed with the rest of your code without encountering issues related to non-numeric time series objects.

[kinantytasyao]

Halo @princessblackswan , karena tidak ada konfirmasi lebih lanjut dan untuk menjaga agar forum diskusi tetap berjalan kondusif, kami izin melakukan lock/close diskusi ini. Semoga dapat dipahami.