In the current version of code,
Nw=10; ax.plot(R[Nw,Nw:-1])
is used to exhibit the changpoints. Although it works fine, I am really confused about the moral behind it. I tried to plot the run length with maximum prob in each time step i.e. the y index of maximum prob in each x col, but the result showed the run length keeps going up... I also went back to Admas's paper but found nothing about change point indentification stuff (he just stop at R matrix)... I also tried to find Adams's MATLAB code, but the code seems to have been removed...

I am trying to use this method in my work, and I believe it's the best to fully understand it before any deployment. Any help will be appreciated and thanks a lot!

Even the input is like In [none]. Something going wrong?

Hi. I was wondering if you had any insight in extending your code to include other emission models besides gaussian. In particular, how about a GMM with known number of gaussians?

I was going to take a stab at implementing it and submit a PR, but wanted to get your input first.

Thanks

Dan

If I click on the "example notebook" work - an nbviewer link - I get a "too many redirects" error.

It would be nice if the example notebook was easily accessible in the repo (maybe I overlooked it... ) because we don't need a live notebook / nbviewer to figure out whether the example fits our use case.

There is always a tradeoff between false alarms and missed alarms, and when the algorithm is more sensitive we should have higher false alarm rate and lower missed alarm rate. My question is, is it possible to adjust the sensitivity level of this algorithm by changing the hyperparameter (e.g., alpha, beta, kappa, mu)? Thank you!

Hi，nice to meet you，and i want to aks a basic question，if i don’t know the distribution of data（not the normal distribution），then could i use the bocpd？
Thank you！

The API in the current state is pretty much tailored to my needs. Would be interesting to make it more general usable. Any ideas anyone?

Hi,

I installed bayesian_changepoint_detection from this github repository.

By setting (accidentally) np.seterr(all='raise'), I was able to cause the following exception.

I am not sure whether this would have any relevance for the further processing, but I just wanted to draw attention to people working on / with this library.

/home/user/venv/env01/bin/python3.6 /home/user/PycharmProjects/project01/snippet.py
Use scipy logsumexp().
Traceback (most recent call last):
  File "/home/user/PycharmProjects/project01/snippet.py", line 68, in <module>
    Q, P, Pcp = offcd.offline_changepoint_detection(data, partial(offcd.const_prior, l=(len(data) + 1)), offcd.gaussian_obs_log_likelihood, truncate=-40)
  File "/home/user/experiments/original-unforked/bayesian_changepoint_detection/bayesian_changepoint_detection/offline_changepoint_detection.py", line 98, in offline_changepoint_detection
    Q[t] = np.logaddexp(P_next_cp, P[t, n-1] + antiG)
FloatingPointError: underflow encountered in logaddexp

Process finished with exit code 1

Hi

awesome package, thanks for putting this together. Quick question: once you have the change point, have you also thought about the direction from that change point?

so far, I have tried:

1. Comparing to data around prev change points: the issue with this is that it might not have identified all change points
2. Complaring data before/after the change point. the change point detection could be delayed, so, not sure how many points to go back. Also, if I check the change point with the next couple/few points, it slows me down

ta!

Please correct me if I am wrong.

It seems like the R matrix is conflating the joint probability and the conditional probability. The last step in the iteration turns it into a conditional probability. When you then calculate the growth/changepoint probabilities, you are using the conditional probabilities instead of the joint probabilities as depicted in the paper.

I think my question is related to the one, which was not answered and is already closed:
#19

In your example, you have applied the student t-distribution as a likelihood. I understand the distribution, its parameters, but I have a question about how you set up prior and update its parameters in the code. So the following is:

df = 2*self.alpha
scale = np.sqrt(self.beta * (self.kappa+1) / (self.alpha * self.kappa))

I don't understand what alpha, beta and kappa correspond to. How have you come across this expression? The paper by Adams and McKey refers to updating sufficient statistics. Is your expression related to that? If so, how can I do that for any other distribution, let's say gaussian? In my comment, I refer to the following formula in the paper:

When we find the p(r_0 | x_{1:t}) you have

R[0, t+1] = np.sum( R[0:t+1, t] * predprobs * H)

but shouldn't it be

R[0, t+1] = np.sum( R[0:t+1, t] * predprobs[0] * H)

as we calculate the predictive probability assuming no data.

Hi,
I've noticed is the scaling of the data can have an effect on the result, but I am not sure why it would and can't find any reason for it in the code or references. Below I have the CP probabilities for the same data with or without a constant factor, which are somewhat different.

Are there some assumptions about the input data I am missing?
Thanks

Was wondering if you had any plans or need for the online extension seen in "Adaptive Sequential Bayesian Change Point Detection."

I was going to implement it for a work project, and would like to just keep it in this package as it fits the theme.

Thoughts?

Was removed during a PR. Is there a good way to keep cython and python in sync. I'm not sure if I prefer one over the other (python is better for debugging, cython is faster).

I read the paper, but I didn't find the reason why the parameters is updated in that way.
Besides, I am thinking about replacing the student_T with gamma, but I don't know how to update the parameters. Can you give me some advice?

Thank you!

line 4 of offline_changepoint_detection.py is importing comb from scipy.misc. it's been deprecated and now it looks like you have to import comb from scipy.special.

I ran the entire sample code and got an error on the log function for online detection. It says "divide by zero encountered in log". I constantly got this error for the random time series or my own time series data. I am wondering why there are zeros and how to avoid them. Could you please take a look when you get some time? Thank you so much!

Hi,

I am confused about the returned R matrix interpretation in the online detection algorithm. In the notebook example, the third plot is R[Nw,Nw:-1], where it is mentioned to be "the probability at each time step for a sequence length of 0, i.e. the probability of the current time step to be a changepoint."
So why do we choose the indices R[Nw,Nw:-1] ? why not R[Nw,:]

Also, it was mentioned as an example that R[7,3] means the probability at time step 7 taking a sequence of length 3, so does R[Nw,Nw:-1] means that we are taking all the probabilities at time step Nw ?

Any suggestions to help me to understand the output R ?

Thanks

Hi. According to the paper, predprobs is defined as the probability of x_t givn r_(t-1) and x_t^r. To my knowledge, it means at each time t, we should compute the predictive distribution given different x_t^r. if r_t = r_(t-1) +1 , let's assume r_t = 4, t =10. we can use x9,x8,x7,x6,x5 to calculate the posterior probablity and thus the predictive distribution. But based on predprobs = observation_likelihood.pdf(x) , the predprobs[i] is the predictive distribution after observing the first i data points. It doesn't seem to be consistent with the original definition. But it works well. I have been struggling with this issue for a while.
Correct me if I misunderstand the concept. Thanks.

Hello,
do you have any idea about online multivariate change point detection?

The algorithm runs online, but with the assumption we have the length of the dataset a priori. What about streaming scenarios where we have a continuous stream of data? Is there (an efficient) way to the run the online algorithm without knowing the length of the dataset?

For someone whoever is interested, I have added Normal known precision, poisson distributions in my fork below. Also tried adding another type of hazard function which is normally distributed over time. Usage of the same is updated in Example code as well. Find my fork here - https://github.com/kmsravindra/bayesian_changepoint_detection

Hello. First of all thank you for such a nice package. Wanted to ask you if the online change point detection has option to train on GPU. I went through the doc but couldn't find anything myself.

Thank you in advance :)

Would be interesting to have champ implemented. See http://www.andrew.cmu.edu/user/sniekum/pubs/CPD15.pdf

Hi guys,

there is an import issue if one uses newer scipy versions.

Would be a quick fix if you adapt the import statement at offline_changepoint_detection.py

try:  # SciPy >= 0.19
    from scipy.special import comb, logsumexp
except ImportError:
    from scipy.misc import comb, logsumexp  # noqa

Is it possible to make Ryan Adams algorithm to work on multivariate data too?

I ran the given online detection example in the notebook, and I assumed the y axis indicating the probability of changepoint (am I right?). But the y value ranged from zero to hundreds.
I am not very familiar with the math, so can anyone please explain this outcome?

Thanks.

Is it possible to get a new release in PyPi? since the offline detection does not work in the release currently in PyPi.

Thank you