Framework for Scalable Dictionary learning for Time Series (SDTS) training function

This function trains a model that uses a dictionary to predict state changes. Differently from fluss(), it doesn't look for semantic changes (that may be several), but for binary states like "on" or "off". Think for example that a human annotator is pressing a switch any time he thinks that the recorded data is relevant, and releases the switch when he thinks the data is noise. This algorithm will learn the switching points (even better) and try to predict using new data.

sdts_train(
  data,
  label,
  window_size,
  beta = 1,
  pat_max = Inf,
  parallel = FALSE,
  verbose = getOption("tsmp.verbose", 2)
)

Arguments

data	a `vector` of `numeric`. Time series.
label	a `vector` of `logical`. Annotations.
window_size	an `int` or a `vector` of `int`. Sliding window sizes.
beta	a `numeric`. See details. (default is `1`).
pat_max	an `int`. Max number of shape features captured. (default is `Inf`).
parallel	a `logical`. Use parallel computation inside (default is `FALSE`).
verbose	an `int`. See details. (Default is `2`).

Value

Returns a list with the learned dictionary score (estimated score), score_hist (history of scores), pattern (shape features), thold (threshold values).

Details

beta is used to balance F-score towards recall (>1) or precision (<1). verbose changes how much information is printed by this function; 0 means nothing, 1 means text, 2 adds the progress bar, 3 adds the finish sound.

References

Yeh C-CM, Kavantzas N, Keogh E. Matrix profile IV: Using Weakly Labeled Time Series to Predict Outcomes. Proc VLDB Endow. 2017 Aug 1;10(12):1802-12.

Website: https://sites.google.com/view/weaklylabeled

Examples

# This is a fast toy example and results are useless. For a complete result, run the code inside
#' Not run' section below.
w <- c(110, 220)
subs <- 11000:20000
tr_data <- mp_test_data$train$data[subs]
tr_label <- mp_test_data$train$label[subs]
te_data <- mp_test_data$test$data[subs]
te_label <- mp_test_data$test$label[subs]
model <- sdts_train(tr_data, tr_label, w, verbose = 0)
predict <- sdts_predict(model, te_data, round(mean(w)))
sdts_score(predict, te_label, 1)
if (FALSE) {
windows <- c(110, 220, 330)
model <- sdts_train(mp_test_data$train$data, mp_test_data$train$label, windows)
predict <- sdts_predict(model, mp_test_data$test$data, round(mean(windows)))
sdts_score(predict, mp_test_data$test$label, 1)
}