pytsmp 0.3.0

A Python implementation of the matrix profile

pytsmp is a Python implementation of the matrix profile. More details about matrix profile can be found in the UCR Matrix Profile Page by the paper authors.

Currently support MASS and the matrix profile algorithms STAMP, STOMP, SCRIMP++ (no multi-core or GPU support yet), and some convenience functions such as discords and motifs finding. I plan to implement the parallelized version of the matrix profile algorithms later.

The original implementation (in R) of the paper authors from the UCR group can be found here.

Installation

pytsmp is available via pip.

pip install pytsmp

Usage

To compute the matrix profile using STAMP, use the following code.

import numpy as np
from pytsmp import STAMP

# create a 1000 step random walk and a random query
ts = np.cumsum(np.random.randint(2, size=(1000,)) * 2 - 1)
query = np.random.rand(200)

# Create the STAMP object. Note that computation starts immediately.
mp = STAMP(ts, query, window_size=50)  # window_size must be specified as a named argument

# get the matrix profile and the profile indexes
mat_profile, ind_profile = mp.get_profiles()

Incremental of the time series and the query is supported.

import numpy as np
from pytsmp import STAMP

# create a 1000 step random walk and its matrix profile
ts = np.cumsum(np.random.randint(2, size=(1000,)) * 2 - 1)
mp = STAMP(ts, window_size=50)
mat_profile, _ = mp.get_profiles()

# create the matrix profile of the first 999 steps and increment the last step later
mp_inc = STAMP(ts[:-1], window_size=50)
mp_inc.update_ts1(ts[-1])  # similarly, you can update the query by update_ts2()
mat_profile_inc, _ = mp_inc.get_profiles()

print(np.allclose(mat_profile, mat_profile_inc))  # True

Benchmark

Perform a simple trial run on a random walk with 40000 data points.

import numpy as np
from pytsmp import STAMP

np.random.seed(42)  # fix a seed to control randomness
ts = np.cumsum(np.random.randint(2, size=(40000,)) * 2 - 1)

# ipython magic command
%timeit mp = STAMP(ts, window_size=1000, verbose=False, seed=42)

# and similarly for STOMP and SCRIMP

On my MacBook Pro with 2.2 GHz Intel Core i7, the results are (all over 7 runs, 1 loop each):

Algorithm	Data Size	Window Size	Elapsed Time
STAMP	40000	1000	2min 14s ± 392ms
STOMP	40000	1000	22.1s ± 52.8ms
SCRIMP (without PreSCRIMP)	40000	1000	23.6s ± 402ms
PreSCRIMP (Approximate algorithm)	40000	1000	606ms ± 9.5ms

Reference

C.C.M. Yeh, Y. Zhu, L. Ulanova, N. Begum, Y. Ding, H.A. Dau, D. Silva, A. Mueen and E. Keogh. “Matrix profile I: All pairs similarity joins for time series: A unifying view that includes motifs, discords and shapelets”. IEEE ICDM 2016.

Y. Zhu, Z. Zimmerman, N.S. Senobari, C.C.M. Yeh, G. Funning, A. Mueen, P. Berisk and E. Keogh. “Matrix Profile II: Exploiting a Novel Algorithm and GPUs to Break the One Hundred Million Barrier for Time Series Motifs and Joins”. IEEE ICDM 2016.

Y. Zhu, C.C.M. Yeh, Z. Zimmerman, K. Kamgar and E. Keogh. “Matrix Profile XI: SCRIMP++: Time Series Motif Discovery at Interactive Speed”. IEEE ICDM 2018.

Disclaimer

This project is for my own learning and understanding purpose, and I may not be able to actively develop it from time to time. If you need a Python implementation of the matrix profile, you may try matrixprofile-ts.