Feature request: Cost function equivalent to R changepoint library cpt_meanvar

[drandynisbet]

Hi,
apologies in advance if I've missed something obvious, I'm using ruptures more as a black-box user - and have only started to dabble with changepoint analysis relatively recently. Could you outline what's required to implement, or whether you might release support for a cost-function covering both mean and variance that would give similar results to the R cpt_meanvar?
I'm using PELT to examine some observations from tracing the performance of applications when they are in execution.

Thanks,
Andy

[oboulant]

Hi @drandynisbet ,

Thanks for you interest in ruptures !

Why not checking for instance Cost Normal that do detect changes in the mean and covariance matrix of a sequence of multivariate Gaussian random variables ?

For instance :

import ruptures as rpt
import numpy as np
from numpy.random import RandomState
import matplotlib.pyplot as plt  # for display purposes

rd = RandomState(1234)
segment_length = 100

signal = np.r_[rd.normal(0, 0.1, segment_length), 
        rd.normal(0, 0.01, segment_length),
        rd.normal(1, 0.05, segment_length)]


algo = rpt.Pelt(model="normal", min_size=1, jump=1).fit(signal)
my_bkps = algo.predict(pen=100)
print(my_bkps)

fig, ax = rpt.display(signal, [100, 200, 300], my_bkps)
plt.show()

Which outputs :

[deepcharles]

As mentionned by @oboulant, this cost function will simultaneously detect changes in the mean and variance/covariance matrix for multidimensional signals. This should emulate the behaviour of cpt_meanvar.

[oboulant]

I am closing the issue. Feel free re-open one or a discussion thread if needed !