In franzbischoff/tsmp: Time Series with Matrix Profile

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-"
)
library(tsmp)

Time Series with Matrix Profile

| | Build | Dev | |--------------|-------|-----| | Linux x86_64 | | | | OSX | | | | Windows | | | | Coverage | | |

Overview

R Functions implementing UCR Matrix Profile Algorithm (http://www.cs.ucr.edu/~eamonn/MatrixProfile.html).

This package allows you to use the Matrix Profile concept as a toolkit.

This package provides:

• Algorithms to build a Matrix Profile: STAMP, STOMP, SCRIMP++, SIMPLE, MSTOMP and VALMOD.
• Algorithms for MOTIF search for Unidimensional and Multidimensional Matrix Profiles.
• Algorithm for Chains search for Unidimensional Matrix Profile.
• Algorithms for Semantic Segmentation (FLUSS) and Weakly Labeled data (SDTS).
• Algorithm for Salient Subsections detection allowing MDS plotting.
• Basic plotting for all outputs generated here.
• Sequencial workflow, see below.

# Basic workflow:
matrix <- tsmp(data, window_size = 30) %>%
  find_motif(n_motifs = 3) %T>%
  plot()

# SDTS still have a unique way to work:
model <- sdts_train(data, labels, windows)
result <- sdts_predict(model, data, round(mean(windows)))

Please refer to the User Manual for more details.

Please be welcome to suggest improvements.

Performance on an Intel(R) Core(TM) i7-7700 CPU @ 3.60GHz using a random walk dataset

set.seed(2018)
data <- cumsum(sample(c(-1, 1), 40000, TRUE))

Current version benchmark

w <- 1000
workers <- 8
data_size <- 40000
set.seed(2018)
times <- 5L
mbm_stomp <- round(median(microbenchmark::microbenchmark(tsmp(data, window_size = w, mode = "stomp", verbose = 0), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)
mbm_stomp_par <- round(median(microbenchmark::microbenchmark(tsmp(data, window_size = w, mode = "stomp", n_workers = workers, verbose = 0), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)
mbm_scrimp <- round(median(microbenchmark::microbenchmark(tsmp(data, window_size = w, mode = "scrimp", verbose = 0), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)
times <- 10L
mbm_mpx <- round(median(microbenchmark::microbenchmark(mpx(data, window_size = w), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)
mbm_mpx_par <- round(median(microbenchmark::microbenchmark(mpx(data, window_size = w, n_workers = workers), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)
times <- 3L
mbm_stamp <- round(median(microbenchmark::microbenchmark(tsmp(data, window_size = w, mode = "stamp", verbose = 0), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)
mbm_stamp_par <- round(median(microbenchmark::microbenchmark(tsmp(data, window_size = w, mode = "stamp", n_workers = workers, verbose = 0), times = times, setup = (data <- cumsum(sample(c(-1, 1), data_size, TRUE))))$time) / 10^9, 2)

bench_data <- data.frame("Elapsed Time(s)" = c(mbm_stamp, mbm_stamp_par, mbm_stomp, mbm_stomp_par, mbm_scrimp, mbm_mpx, mbm_mpx_par), "Data Size" = data_size, "Window Size" = w, Threads = c(1, workers, 1, workers, 1, 1, workers), Lang = c("R", "R", "R", "R", "R", "Rcpp", "Rcpp"), row.names = c("`stamp`", "`stamp_par`", "`stomp`", "`stomp_par`", "`scrimp`", "`mpx`", "`mpx_par`"), check.names = FALSE)

knitr::kable(bench_data[order(bench_data$`Elapsed Time(s)`),])

Installation

# Install the released version from CRAN
install.packages("tsmp")

# Or the development version from GitHub:
# install.packages("devtools")
devtools::install_github("matrix-profile-foundation/tsmp")

Currently available Features

• STAMP (single and multi-thread versions)
• STOMP (single and multi-thread versions)
• STOMPi (On-line version)
• SCRIMP (single-thread, not for AB-joins yet)
• Time Series Chains
• Multivariate STOMP (mSTOMP)
• Multivariate MOTIF Search (from mSTOMP)
• Salient Subsequences search for Multidimensional Space
• Scalable Dictionary learning for Time Series (SDTS) prediction
• FLUSS (Fast Low-cost Unipotent Semantic Segmentation)
• FLOSS (Fast Low-cost On-line Unipotent Semantic Segmentation)
• SiMPle-Fast (Fast Similarity Matrix Profile for Music Analysis and Exploration)
• Annotation vectors (e.g., Stop-word MOTIF bias, Actionability bias)
• FLUSS Arc Plot and SiMPle Arc Plot
• Exact Detection of Variable Length Motifs (VALMOD)
• MPdist: Matrix Profile Distance
• Time Series Snippets
• Subsetting Matrix Profiles (head(), tail(), [, etc.)
• Misc:
• MASS v2.0
• MASS v3.0
• MASS extensions: ADP (Approximate Distance Profile, with PAA)
• MASS extensions: WQ (Weighted Query)
• MASS extensions: QwG (Query with Gap)
• Fast moving average
• Fast moving SD

Roadmap

• Profile-Based Shapelet Discovery
• GPU-STOMP

Other projects with Matrix Profile

• Python: https://github.com/target/matrixprofile-ts
• Python: https://github.com/ZiyaoWei/pyMatrixProfile
• Python: https://github.com/jbeleno/owlpy
• Python: https://github.com/javidlakha/matrix-profile
• Python: https://github.com/shapelets/khiva-python
• R: https://github.com/shapelets/khiva-r
• Matlab: https://github.com/shapelets/khiva-matlab
• Java: https://github.com/shapelets/khiva-java
• Java: https://github.com/ensozos/Matrix-Profile
• Kotlin: https://github.com/shapelets/khiva-kotlin
• C++ (CUDA and OPENCL): https://github.com/shapelets/khiva
• CUDA: https://github.com/zpzim/STOMPSelfJoin
• CUDA: https://github.com/zpzim/SCAMP

Matrix Profile Foundation

Our next step unifying the Matrix Profile implementation in several programming languages.

Visit: Matrix Profile Foundation

Code of Conduct

Please note that the 'tsmp' project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

franzbischoff/tsmp documentation built on March 9, 2020, 6:01 a.m.