I note in the paper you talk about different penalties in section 6.1. However from looking through the library, it seems that ruptures only supports a fixed linear penalty (ie Beta). Am I right to assume that it doesn't work with more complex penalties linear such as AIC?
Further, if I wanted to implement a method were you normally would calculate a p-value of splitting (ie a likelihood ratio test following Chi squared), is the idea that we just have the error() method return the test statistic without testing for significance (ie the raw likelihood ratio), and the penalty constant implies a p-value? I suppose this makes segmentation fast and flexible, but highly dependent on the choice of penalty.
I note in the paper you talk about different penalties in section 6.1. However from looking through the library, it seems that ruptures only supports a fixed linear penalty (ie Beta). Am I right to assume that it doesn't work with more complex penalties linear such as AIC?
Further, if I wanted to implement a method were you normally would calculate a p-value of splitting (ie a likelihood ratio test following Chi squared), is the idea that we just have the
error()
method return the test statistic without testing for significance (ie the raw likelihood ratio), and the penalty constant implies a p-value? I suppose this makes segmentation fast and flexible, but highly dependent on the choice of penalty.