attic/rfc_window_conversion.md

In mlr-org/mlr3forecasting: Forecasting for 'mlr3'

Feature Name: `Windowconversion`
Start Date: 2019-12-04
Target Date:

Summary

• function prototype for a windowconversion using lagged and lagged differenced values

Motivation

To solve forecasting problems as a regression problems, reshaping the data to an appropriate format is essential. Supervised ML-Algorithms usually expect the data to be in a tabular format: one row euqals one observation. The transformation to such a structure is called window-conversion.

Guide-level explanation

Example for a univariate forecasting problem

Consider the following timeseries-data:

data
        date target 
1 2019-11-29      1 
2 2019-11-30      2
3 2019-12-01      4
4 2019-12-02      4
5 2019-12-03      9
6 2019-12-04      6
.      .          .
.      .          .
.      .          .

One approach to forecast this series could be, to use two lagged values and the first difference of the series as features. (For the differences see https://www.rdocumentation.org/packages/base/versions/3.6.1/topics/diff )

The window conversion should be a function that creates additional columns for every lag and every lagged diffference we specified.

windowConversion(data, date.col="date", lag = c(1L,2L), diff.lag = c(1L), diff.diff = c(1L))

        date target target_lag_1 target_lag_2 target_lag_1_diff_1
1 2019-11-29      1           NA           NA         NA
2 2019-11-30      2            1           NA          1
3 2019-12-01      4            2            1          2
4 2019-12-02      4            4            2          0
5 2019-12-03      9            4            4          5
6 2019-12-04      6            9            4         -3
.      .          .            .            .          .
.      .          .            .            .          .
.      .          .            .            .          .

Reference-level explanation

A prototype of the function for the window-conversion including lagged and lagged differenced features. This may be used as a basis for discussions.

windowConversion = function(data, date.col = NULL, lag.cols = NULL, lag = 0L,
                            diff.cols=NULL, diff.lag=0L, diff.diff=0L){

  #... include some checks for dataformat

  setDT(data)
  if(is.null(date.col)){
    stop("date.col must be given")
  }
  assert_choice(date.col, colnames(data))
  assert_integerish(lag)
  assert_integerish(diff.lag)
  assert_integerish(diff.diff)
  assert_subset(lag.cols, colnames(data))
  assert_subset(diff.cols, colnames(data))
  # if not specified use all columns except the date column
  if(is.null(diff.cols)){
    diff.cols = setdiff((colnames(data)),date.col)
  }
  if(is.null(lag.cols)){
    lag.cols = setdiff((colnames(data)),date.col)
  }

  if(any(lag>0)){
    lag.names = do.call("paste", c(CJ(lag.cols,"lag", lag, sorted = FALSE), sep="_"))
    data[, c(lag.names) := shift(.SD, lag), .SDcols = lag.cols]
  }

  if(any(diff.lag)>0 | any(diff.diff>0)){
    #diff only accepts values > 0
    if(!any(diff.lag > 0)){
      diff.lag = 1
    }
    if(!any(diff.diff > 0)){
      diff.diff = 1
    }

    # helper function to adjust for the shorter output of diff() with NA padding
    padDiff = function(x, lag,diff){
      xx = diff(x,lag,diff)
      pad = rep(NA,length(x)-length(xx) )
      return(c(pad,xx))
    }

    #Combinations of lags and differences
    diff.selection = CJ(diff.lag,diff.diff)

    for(i in seq_len(nrow(diff.selection))){
      lag.val = diff.selection[[i,1]]
      diff.val = diff.selection[[i,2]]
      lagdiff.names = do.call("paste", c(CJ(diff.cols,"lag", lag.val,"diff",diff.val, sorted = FALSE),sep="_"))
      data[,c(lagdiff.names) := lapply(.SD,padDiff,lag.val,diff.val)  ,.SDcols = diff.cols ]
    }
  }
  return(data)
}

Rationale, drawbacks and alternatives

data.table currently has no function like diff() in base R, which makes getting the lagged differences kind of ugly.

Prior art

Introduced Dependencies

The above function relies on the following packages:

• checkmate
• data.table

Unresolved questions

mlr-org/mlr3forecasting documentation built on June 29, 2023, 11:57 p.m.