Check in a series if there exists adjacent values with less than a certain number of missing values

Question

I'm working with a large number of series of different experiments, where there exist a lot of experiments where there are no entries. I now want to draw a sample where there should exist at least 200 adjacent values of data with less than 100 entries of null (A 200 day period of complete data).

I started to do it with a loop, checking for 0-200, 1-200 for each dataset which takes ages... Should I use Row / Scan/ Reap/ Sow? I'd appreciate a help and guidance on possible functions I can use.

Answer 1

Here is a rule-based version, which may be not the fastest but seems elegant (this is based on the same idea as the solution described in this answer):

Clear[getSamples];
getSamples[data_List, minlen_Integer?Positive, n : (_Integer | All) : All] :=
   ReplaceList[
     data,
     {x___, p : Repeated[Except[0 | Null], {minlen, Infinity}], ___} :> {{p}, Length[{x}+1]},
     If[n === All, Sequence @@ {}, n]
   ]

It will return some number (ot all) of samples together with their starting positions. Here is a test sample (much smaller than what you need, just for an illustration):

testSample = RandomChoice[{0, Null}~Join~Range[5], 30]

(*
   ==>  {0, 3, 1, 0, 5, 5, 1, 2, Null, 1, 2, 2, 1, 0, 4, 1, 4, 3, 0, 0, 4, 1, 4, 0, 5,
           Null, 1, 4, Null, 2}
*)

Here are examples. First, find all samples with length at least 4:

getSamples[testSample, 4]

(*
   ==> {{{5, 5, 1, 2}, 5}, {{1, 2, 2, 1}, 10}, {{4, 1, 4, 3}, 15}}
*)

And now just the first of them (faster):

getSamples[testSample, 4, 1]

(*
   ==>  {{{5, 5, 1, 2}, 5}}
*)

This assumes that you have your data in the format of testSample.

An alternative way would be to use Partition, but it may be memory - inefficient for very large data (you create a list occupying about 250 times the memory of the original one, for your case). It can also be slower, if the needed sample can be found very early on, although typically will be faster. Still, here is the code:

Clear[getSamplesP];
getSamplesP[data_List, minlen_Integer?Positive, n : (_Integer | All) : All] := 
   Select[
     Transpose[{#, Range[Length[#]]}] &@ Partition[data, minlen, 1], 
     ! MemberQ[First@#, 0 | Null] &, 
     If[n =!= All, n, Sequence @@ {}]
]

you can use this in the same way.

Answer 2

At Leonid's prodding, and some mulling over on how to modify my original idea, I think this does what you want:

getSamples[data_List, minlen_Integer?Positive, n:(_Integer|All):All] :=
Module[{spl = SplitBy[data, MatchQ[#, 0|0.|Null] &]},
       Take[Select[Transpose[{spl, (Accumulate[#] - # + 1) &[Length /@ spl]}], 
                   (Length[First[#]] >= minlen) &], n]]

The SplitBy[] is key here; data is split according to whether it contains 0, 0. or Null (that is, with the pattern 0|0.|Null), and the sublists of length minlen or longer are taken, I'll use Leonid's example here as well:

testSample = {0, 3, 1, 0, 5, 5, 1, 2, Null, 1, 2, 2, 1, 0, 4, 1, 4,
              3, 0, 0, 4, 1, 4, 0, 5, Null, 1, 4, Null, 2};

getSamples[testSample, 4]
{{{5, 5, 1, 2}, 5}, {{1, 2, 2, 1}, 10}, {{4, 1, 4, 3}, 15}}

getSamples[testSample, 4, -2]
{{{1, 2, 2, 1}, 10}, {{4, 1, 4, 3}, 15}}

Positive n does what you expect; negative values of n, as with Mathematica convention, take from the right.

---

If one wants to account for overlapping as well, some modification is necessary (borrowing slightly from Leonid):

getSamples[data_List, minlen_Integer?Positive, n:(_Integer|All):All] := 
 Module[{da = SplitBy[data, MatchQ[#, 0|0.|Null] &]},
  da = Function[l, 
     Transpose[
      Through[{Identity, Last[l] + Range[Length[#]] - 1 &}[
        Partition[First[l], minlen, 1]]]]] /@ 
    Select[Transpose[{da, (Accumulate[#] - # + 1) &[
        Length /@ da]}], (Length[First[#]] >= minlen) &];
  Take[Flatten[da, 1], n]]

As an example:

getSamples[testSample, 3]
{{{5, 5, 1}, 5}, {{5, 1, 2}, 6}, {{1, 2, 2}, 10}, {{2, 2, 1}, 11},
 {{4, 1, 4}, 15}, {{1, 4, 3}, 16}, {{4, 1, 4}, 21}}

Answer 3

Suppose $p_k$ with $1\leq k\leq N_0$ is the position of the $k$-th zero in the data list $\{a_1,\ldots,a_N\}$, then each sublist which contains $\leq M$ zeros will be a sublist of one of the sublists $$ \{ a_{p_k+1},\ldots, a_{p_{k+M+2}-1}\}, \qquad 1\leq k \leq N_{0}-M-2 $$ or $\{ 1,\ldots, a_{p_{M+1}-1}\}$ or $\{a_{p_{N_0-M-1}+1},\ldots, a_N\}$, so in Mathematica speak it is sufficient to construct the lists

Take[lst {zlst[[k]]+1 ;; zlst[[k+maxZeros+2]]-1}]

for which zlst[[k+maxZeros+2]]-zlst[[k]]-1 >= minLen where zlst is the list of positions of the zeros in lst padded with 0 on the left and Length[lst]+1 on the right. This could be implemented as

findLsts2[lst_, minLen_Integer?Positive, maxZeros_, 
  n : (_Integer | All) : All] :=
 Module[{zeroLst, rangeLst},
  zeroLst = Flatten[Position[lst, (0 | 0. | Null)]];
  If[Length[zeroLst] <= maxZeros,
   lst,
   rangeLst = Transpose[{Prepend[Drop[zeroLst + 1, -maxZeros], 1],
      Append[Drop[zeroLst - 1, maxZeros], Length[lst]]}];
   {Take[lst, #], #} & /@ 
    Select[rangeLst, (#[[2]] - #[[1]] + 1) >= minLen &, 
     n /. All -> Sequence[]]
   ]
  ]

This would return the sublists of maximum length which contain at most maxZeros zeros together with the indices of these sublists in the data list lst. Example

findLsts2[{1, 0, 2, 0, 3, 0, 0, 0, 2, 4, 5, 6, 7, 0}, 4, 2] // MatrixForm

Edit

This updated version of findLsts2 randomly selects n sublists from all possible sublists of length len with at most maxZeros zeros from lst.

findLsts3[lst_, len_Integer?Positive, maxZeros_, 
  n : (_Integer | All) : All] :=
 Module[{zeroLst, rangeLst, startLst},
  zeroLst = Flatten[Position[lst, (0 | 0. | Null)]];
  If[Length[zeroLst] <= maxZeros,
   startLst = Range[Length[lst] - len + 1],
   rangeLst = Transpose[{Prepend[Drop[zeroLst + 1, -maxZeros], 1],
      Append[Drop[zeroLst - len, maxZeros], Length[lst] - len + 1]}];
   startLst = Flatten[Range @@@ (List @@ (Interval @@ 
          Select[rangeLst, (#[[2]] - #[[1]]) >= 0 &])), 1]];
  {lst[[# ;; # + len - 1]], #} & /@ Sort@RandomSample[startLst, n]
  ]

For the previous example we get

findLsts2[{1, 0, 2, 0, 3, 0, 0, 0, 2, 4, 5, 6, 7, 0}, 4, 2] // MatrixForm

Edit 2

The $k$-th pair in rangelst in findLsts3 corresponds to the values $\{j_1, j_2\}$ such that $$p_{k-1}+1 = j_1 \qquad\text{and}\qquad j_2+l-1 = p_{k+M+1}-1$$ where $l$ is the length of the sublist you want to extract and $p_k$ is as above. Provided that $j_{1}\leq j_{2}$ we then have that for all $j_{1}\leq j\leq j_{2}$ the sublist $\{a_j,\ldots,a_{j+l-1}\}$ is a subset of $\{ a_{p_{k-1}+1},\ldots, a_{p_{k+M+1}-1}\}$. What the Select statement in startLst does is to filter out the pairs in rangelst for which $j_1\leq j_2$ doesn't hold. startLst is then the Union of the ranges Range[j1, j2] for all remaining pairs.

Answer 4

A simple solution uses Partition and Count:

$data = {1,0,2,0,3,0,Null,0,2,4,5,6,7,0};
$runLength = 4;
$tolerance = 2;

Cases[Partition[$data, $runLength, 1], l_ /; Count[l, 0|Null] <= $tolerance]

{{1,0,2,0},{0,2,0,3},{2,0,3,0},{Null,0,2,4},{0,2,4,5},{2,4,5,6},{4,5,6,7},{5,6,7,0}}

Unfortunately, this solution does not scale well. It uses a lot of memory by creating and scanning a list containing every run. For large input lists, this is not feasible. Furthermore, it performs a number of operations proportional to the product of the input and run lengths. The following function performs better, using far less memory and with performance linearly proportional to the input size:

goodRunPositions[data_, runLength_, tolerance_] :=
  data /. Null -> 0 //
  Sign //
  Abs //
  1 - # & //
  Accumulate //
  Prepend[#, 0] &//
  #[[runLength+1;;]] - #[[;;-runLength-1]] &//
  Position[#, n_ /; n <= tolerance] &//
  Flatten

goodRunPositions[$data, $runLength, $tolerance]

{1,2,3,7,8,9,10,11}

A helper function can be used to extract a run at any given position:

extractRun[data_, runLength_, position_] :=
  data[[position ;; position+runLength-1]]

extractRun[$data, $runLength, 7]

{Null,0,2,4}

... or to randomly select a run ...

randomGoodRun[data_, runLength_, tolerance_] :=
  extractRun[
    $data
  , $runLength
  , RandomChoice @ goodRunPositions[data, runLength, tolerance]
  ]

randomGoodRun[$data, $runLength, $tolerance]

{2,4,5,6}

Performance is adequate for larger lists:

randomData[w_, n_] :=
  RandomChoice[{w, w, 100, 100, 100, 100} -> {Null, 0, 1, 2, 3, 4}, n]

SeedRandom[923874]
$10K = randomData[20, 10000];
$1M = randomData[25, 1000000];
$10M = randomData[30, 10000000];

goodRunPositions[$10K, 250, 10] // Timing // Column

0.016 {7425,7426,7427,7428,7429,7430,7431,7432,7433,7434,7435,7436,7437,7438,7439,7440,7441,7442,7443,7446,7447,7448,7449,7450,7451,7452,7453,7454,7455,7456,7457,7458,7459,7460,7461,7462,7463,7464,7465,7466,7467,7468,7469,7470,7471,7472,7473,7474,7475,7476,7477,7478,7479,7480,7481,7482,7483,7484,7485,7486,7487,7488,7489,7490,7491,7492,7493,7494,7495,7496,7497,7498}

goodRunPositions[$1M, 250, 10] // Timing // Column

0.811 {5879,5880,5881,5882,5883,5884,5885,5886,5887,5888,5889,5890,5891,5892,5893,5894,5895,5896,5897,5898,5899,5900,5901,5902,5903,5904,5905,5906,5907,5908,5909,5910,5911,5912,896682,896683,896684}

goodRunPositions[$10M, 250, 10] // Timing // Column

8.221 {3399940,3399941,3399942,3399943,3399944,3399945,3399946,3399947,3399948,3399949,3399950,3399951,3399952,3399953,3399954,3399955,3399956,3399957,3399958,3399959,3399960,3399961,3399962,3399963,3399964,3399965,3399966,3399967,3399968,7027173}

How Does It Work?

Here is our toy data set:

$data

{1,0,2,0,3,0,Null,0,2,4,5,6,7,0}

The first step is to identify nulls and zeros. In order to allow us to use the fast built-in tensor operations, we change nulls to zeroes first:

$data /. Null -> 0 // Sign // Abs // 1 - # &

{0,1,0,1,0,1,1,1,0,0,0,0,0,1}

Every null or zero value has been assigned a score of one and other values are scored zero. We can now generate a running total of the number of zero values encountered:

% // Accumulate

{0,1,1,2,2,3,4,5,5,5,5,5,5,6}

To determine the number of zeroes in any given run, we can simply take the difference between the number of zeroes at the beginning and the end of the run. We'll need to prepend a zero to the list to handle the first run:

Prepend[%, 0] //
#[[$runLength+1;;]] - #[[;;-$runLength-1]] &

{2,2,2,3,3,3,2,1,0,0,1}

Now it is just a simple matter of finding the positions of the elements where the zero count remains below the threshold:

Position[%, n_ /; n <= $tolerance] // Flatten

{1,2,3,7,8,9,10,11}

We can use extractRun inspect the runs themselves:

extractRun[$data, $runLength, #] & /@ % // Column

{1,0,2,0}
{0,2,0,3}
{2,0,3,0}
{Null,0,2,4}
{0,2,4,5}
{2,4,5,6}
{4,5,6,7}
{5,6,7,0}