search

X-DataInitiative / tick

Module for statistical learning, with a particular emphasis on time-dependent modelling
https://x-datainitiative.github.io/tick/
BSD 3-Clause "New" or "Revised" License
484 stars 105 forks source link

The goodness of fit of Hawkes processes (or D-dimentional Hawkes processes) #393

Open rudanie89 opened 5 years ago

rudanie89 commented 5 years ago

Good morning,

I read the paper that tick.HawkesADM4 was developed: 'Learning Social Infectivity in Sparse Low-rank Networks Using Multi-dimensional Hawkes Processes.' For my understanding, this paper focused on optimisation problem, and they use "loglike" metric score to compare with baseline. So, I am wondering that which function from 'tick' provide me to test whether Hawkes processes fit with my data ( I mean the goodness of fit). Thank you.

Mbompr commented 5 years ago

Hi,

Have you tried objective : that is the function to be minimized, or score : that is the goodness of fit function, to test whether Hawkes processes fit with your data but that does not include the penalty terms.

rudanie89 commented 4 years ago

Hi, Thank you. Yes I have tried. But I don't think it is as we usually do in statistics like, we will test how well Hawkes processes can fit with the data. And, I am just wondering that is it weird if I got some coefficient values of adjacency matrix are greater than 1? I though the coefficient values of adjacency matrix are in (0,1). Can you help me to correct me please? Thank you.

stephanegaiffas commented 4 years ago

Hi, what is computed through the score method is precisely the negative log likelihood of the model, so yes it is what we usually do in statistics. However, tick does not provide statistical tests for now, with p-value computations.

The entries of the adjacency matrix can perfectly be larger than 1, note that it is not an adjacency matrix per se from graph theory, but something that plays a similar role for a Hawkes model. The only theoretical constraint about it (but not required in the procedures used in tick) is that it’s spectral radius is < 1 to ensure stationnarity

Best

Le 26 sept. 2019 à 07:31, rudascience notifications@github.com a écrit :

 Hi, Thank you. Yes I have tried. But I don't think it is as we usually do in statistics like, we will test how well Hawkes processes can fit with the data. And, I am just wondering that is it weird if I got some coefficient values of adjacency matrix are greater than 1? I though the coefficient values of adjacency matrix are in (0,1). Can you help me to correct me please? Thank you.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

rudanie89 commented 4 years ago

Hi, Thank you. Yes, I mean p-value. Re: The entries of the adjacency matrix. So, it doesn't mean the influence probabilities of users in HawkesADM4? I thought the adjacency matrix as the following paper mentioned: Zhou, K., Zha, H., & Song, L. (2013, May). Learning Social Infectivity in Sparse Low-rank Networks Using Multi-dimensional Hawkes Processes.

I am so sorry to ask a lot. But, I am confused now. Might you help me to clarify, please? Thank you.