This is a python wrapper for Marteau's implementation of this 2009 paper [1] introducing the Time Warped Edit Distance (TWED, also TWE Distance), slightly modified and provided here. TWED is an elastic distance measure for comparing time series sequences that has been shown to be a promising contender for the popular Dynamic Time Warping (DTW) distance and edit distances that have been adjusted to work on real numbers, such as the Edit Distance on Real Sequences (EDR) and Edit Distance with Real Penalty (ERP) [1, 2]. Like the edit distances, it comprises a penalty for mismatches, and, like DTW, it allows to control for elasticity. In contrast to DTW and the edit distance measures, it is a proper distance metric, as it fulfills the triangle inequality. Like DTW and the edit distances, TWED has a time complexity of O(mn), where m and n are the lengths of the time series sequences compared. Hence, it may not be appropriate for very long sequences and large time series databases. However, the fulfillment of the metric properties allows for using some established speed-up tricks in down-stream algorithms, like nearest neighbor retrieval, that can compensate for the complexity.
The module can be installed from PyPI with
pip install pytwednumpy needs to be installed as the only requirement.
The module can also be installed locally, by checking out this repo or downloading the code and running the following commands in the directory of setup.py
pip install -r requirements.txt
python setup.py build
python setup.py installTo test the installation run the following:
python test/test_pytwed.pyThe example_build.sh provides an example command to build the sources without pip.
The module can then be loaded and used as follows:
import pytwed
import numpy as np
seq1 = np.array([[1.0], [2.1], [3.2], [4.3], [5.4]]).astype('float64')
seq2 = np.array([[9.8], [7.6], [6.5], [4.3], [2.1]]).astype('float64')
ts1 = np.array([1., 2., 3., 4., 5.]).astype('float64')
ts2 = ts1.copy()
pytwed.twed(seq1, seq2, ts1, ts2, nu = 0.001, lmbda = 1.0, p = 2, fast=True)Sequences seq1 and seq2 should be of type numpy.array with shapes (m, d) and (n, d), where m and n are the lengths of the sequences and d is the dimensionality of the sequences. The TWED can be provided with time stamps ts1 and ts2 with type numpy.array. If None is provied (the default), then equally spaced sequences are assumed (they are set to ts1=np.arange(len(seq1)) and ts2=np.arange(len(seq2))).
The distance function has three parameters that can be chosen/tuned. Parameter p refers to the L^p norm that is used for the distance calculation locally for individual elements. For p=2, the Euclidean norm is used, and it is a useful default. Parameter nu (should be > 0, default: 0.001) is a non-negative constant which characterizes the stiffness of the elastic TWED measure. Parameter lmbda (should be >= 0, default: 1.0) is a constant penalty that punishes the editing efforts. The parameters should be tuned for each individual problem and dataset using training data minimizing your task's objective on a hold-out set. Parameter grids to try are for instance nu in [1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1] and lmbda in [0, .25, .5, .75, 1.0] (see [1] for more details).
This package provides two implementations: A fast implementation as a C extension and a pure Python implementation. The Python implementation should only be used as reference implementation. The C implementation is about two orders of magnitude faster. With parameter fast=True the C implementation is used.
[1] Marteau, Pierre-François. "Time warp edit distance with stiffness adjustment for time series matching." IEEE transactions on pattern analysis and machine intelligence 31.2 (2008): 306-318. Url: http://dx.doi.org/10.1109/TPAMI.2008.76).
[2] Serra, Joan, and Josep Ll Arcos. "An empirical evaluation of similarity measures for time series classification." Knowledge-Based Systems 67 (2014): 305-314.