GGS
Greedy Gaussian Segmentation (GGS) is a Python solver for efficiently segmenting multivariate time series data. For implementation details, please see our paper at http://stanford.edu/~boyd/papers/ggs.html.
The GGS Solver takes an n-by-T data matrix and breaks the T timestamps on an n-dimensional vector into segments over which the data is well explained as independent samples from a multivariate Gaussian distribution. It does so by formulating a covariance-regularized maximum likelihood problem and solving it using a greedy heuristic, with full details described in the paper.
Download & Setup
- Download the source code in the terminal by running:
git clone [email protected]:davidhallac/GGS.git
- Confirm that the code was properly downloaded by running:
cd GGS
python helloworld.py
- To write your own Python function that uses ggs, simply make sure that
ggs.py
is in the same directory as your new file, and then add the following code to the beginning of your script:
from ggs import *
Supported Functions
The GGS package has three main functions:
bps, objectives = GGS(data, Kmax, lamb)
Finds K breakpoints in the data for a given regularization parameter lambda
Inputs
data - a n-by-T data matrix, with T timestamps of an n-dimensional vector
Kmax - the number of breakpoints to find
lamb - regularization parameter for the regularized covariance
Returns
bps - List of lists, where element i of the larger list is the set of breakpoints found at K = i in the GGS algorithm
objectives - List of the objective values at each intermediate step (for K = 0 to Kmax)
meancovs = GGSMeanCov(data, breakpoints, lamb)
Finds the means and regularized covariances of each segment, given a set of breakpoints.
Inputs
data - a n-by-T data matrix, with T timestamps of an n-dimensional vector
breakpoints - a list of breakpoint locations
lamb - regularization parameter for the regularized covariance
Returns
meancovs - a list of (mean, covariance) tuples for each segment in the data
cvResults = GGSCrossVal(data, Kmax=25, lambList = [0.1, 1, 10])
Runs 10-fold cross validation, and returns the train and test set likelihood for every (K, lambda) pair up to Kmax
Inputs
data - a n-by-T data matrix, with T timestamps of an n-dimensional vector
Kmax - the maximum number of breakpoints to run GGS on
lambList - a list of regularization parameters to test
Returns
cvResults - list of (lamb, ([TrainLL],[TestLL])) tuples for each regularization parameter in lambList. Here, TrainLL and TestLL are the average per-sample log-likelihood across the 10 folds of cross-validation for all K's from 0 to Kmax
Additional optional parameters (for all three functions above):
features = [] - select a certain subset of columns in the data to operate on
verbose = False - Print intermediate steps when running the algorithm
Example Usage
Running financeExample.py
will yield the following plot, showing the objective (Equation 4 in the paper) vs. the number of breakpoints:
Once we have solved for the locations of the breakpoints, we can use the FindMeanCovs()
function to find the means and covariances of each segment. In the example in helloworld.py
, plotting the means, variances, and covariances of the three signals yields:
To run cross-validation, which can be useful in determining optimal values of K and lambda, we can use the following code to load the data, run the cross-validation, and then plot the test and train likelihood:
from ggs import *
import numpy as np
import matplotlib.pyplot as plt
filename = "Returns.txt"
data = np.genfromtxt(filename,delimiter=' ')
feats = [0,3,7]
#Run cross-validaton up to Kmax = 30, at lambda = 1e-4
maxBreaks = 30
lls = GGSCrossVal(data, Kmax=maxBreaks, lambList = [1e-4], features = feats, verbose = False)
trainLikelihood = lls[0][1][0]
testLikelihood = lls[0][1][1]
plt.plot(range(maxBreaks+1), testLikelihood)
plt.plot(range(maxBreaks+1), trainLikelihood)
plt.legend(['Test LL','Train LL'], loc='best')
plt.show()
The resulting plot looks like:
References
Greedy Gaussian Segmentation of Time Series Data -- D. Hallac, P. Nystrup, and S. Boyd
Authors
David Hallac, Peter Nystrup, and Stephen Boyd.