Description of recovery scenarios

November 30, 2020 · View on GitHub

Setup

all data is z-score normalized
normalization occurs before trimming the length for tests where N_test < N
when percentage or division is mentioned, the result is floored down to nearest integer

Scenarios

N = lentgh of time series M = number of time series

W = 10% * N

I. Batch End

missingpercentage:

N = max; M = max;
size of the missing block varies between 10% and 80% of the series; position: end of 1st series.

length:

M = max; N varies between 20% and 100% of max(N);
size of the missing block = W; position: end of 1st series.

columns:

N = max; M = varies between min(10% of columns, 4) and 100% the number of columns;
size of the missing block = W; position: end of 1st series.

blackout:

N = max; M = max;
Missing data varies between 10 and 100 values of each series; position: end of series.

multicol-increasing:

N = max; M = max;
number of incomplete series varies between 10% and 100% of max(M); size of missing blocks= W; position: at the end of an incomplete time series.

mcar-ts-block*:

N = max; M = max;
number of incomplete series varies between 10% and 100% of max(M); Missing block: 1 per incomplete series, size missing block = 10% * W, position: random within W from the end.

mcar-ts-block2*:

same as ts-block, but only every second incomplete time series.

mcar-ts-multiblock*:

N = max; M = max;
number of incomplete series varies between 10% and 100% of Max(M); sub-blocks of 2% * W are removed at random from each incomplete series until a total of 10% * W of all points in the series are missing. position: random within W from the end.

mcar-ts-multiblock2*:

same as ts-multiblock, but only every second incomplete time series.

mcar-columns*:

N = max; M = varies between max(10% of columns, 4) and 100% of columns;
Missing blocks: 10% * W of every second series total in sub-blocks of 1% * W, start = random within W from the end.

mcar-length*:

M = max; N varies between 20% and 100% of max(N);
Missing blocks: 10% * W of every second series total in sub-blocks of 1% * W, start = random within W from the end.

II. Streaming End