omitakahiro/hawkes

a python package for simulation and inference of Hawkes processes.

hawkes's Issues

number of simulated points - multiple relaizations - bibliografy

Hi,
I discovered today you're nice package for Hawkes process and I'm trying to use it. However, for my purposes, I need to simulate and estimate very long time series, e.g. with a baseline = 1 , branching ratio almost critical = 0.99 and simulation interval of order of 2e6 (or even grater if possible...). I founded that, on simulation interval smaller than the ones I'm interested in and in the case of power law kernel, you're ibrary performs better (in terms of simulationtime) than tick (https://x-datainitiative.github.io/tick/index.html). However, it seems that there's an upper bound on the number of points that can be simulated, 1e6 + 1 I would say. Is there any way to change it?
Moreover, I was wondering whther in both the simulation part and in the estimaiton part you used or not the approximation of the power law kernel introduced in https://link.springer.com/article/10.1140/epjb/e2013-40107-3 .
Finally, I wanted to ask whther is possible to simulate and fit multiple instances from the same system and which papers, more or less, you referred to in the library.
Thank you so much

misc.ipynb is showing some errors while running ?

ModuleNotFoundError: No module named 'Hawkes.BasisFunction'
So I change it to from Hawkes.tools.BasisFunction import loglinear_COS

then it showed
AttributeError: 'loglinear_COS' object has no attribute 'M_BF'

Then I change plot function to plt.plot(x,loglinear.A[:,i],lt[i])

Are these changes correct?

log-linear baseline

Can I understand the log-linear baseline as the co-variates that will influence the occurence of the event?
Then in the patient example(how different kinds of medical activities excite each other and the chance of getting ill), can i describe the heterogeneity between patients with the log linear form of baseline intensity? Thx!

How to install the package?

Dear Author,

I am interested in Hawkes package.
But I don't know how can I use the package, Might you give more detail instructions, please?
Thank you so much.

Best Regards,
Ruda

Multivariate Hawkes Process

I am interested in mutually exciting processes and I was wondering if you have a package for multivariate hawkes process

Overflow encountered in multiply

Hi,
I met a troublesome issue. My event sequence is relatively long (with the length of 6000 that is split into 169 sub-intervals). The program generates a warning:

/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:643: RuntimeWarning: overflow encountered in multiply
  Int = Int + alpha[i] * ( np.exp( -beta[i] * x1 ) - np.exp( -beta[i] * x2 ) )
/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:655: RuntimeWarning: overflow encountered in multiply
  dInt[('alpha',i)] = np.exp( -beta[i] * x1 ) - np.exp( -beta[i] * x2 )
/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:656: RuntimeWarning: overflow encountered in multiply
  dInt[('beta',i) ] = alpha[i] * ( - x1 * np.exp( -beta[i] * x1 ) + x2 * np.exp( -beta[i] * x2 ) )
/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:688: RuntimeWarning: overflow encountered in multiply
  r = np.exp(-beta*step)
/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:699: RuntimeWarning: overflow encountered in double_scalars
  dl = { ('alpha',i):g[i]/alpha[i] for i in range(num_exp) }
/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:709: RuntimeWarning: overflow encountered in add
  g = g + alpha*beta
/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py:690: RuntimeWarning: invalid value encountered in multiply
  g = g*r

It seems the resulting parameters are too large. Could you please give me some suggestions with such a situation?
My codes:

seq = [0, 0.001, 1, ...., 844.001, 844.002]
model = hk.estimator().set_kernel('exp').set_baseline('plinear')
model.fit(seq, (0, 844.1))

I would be really grateful if you could give me some help. Thanks in advance!

BTW, I also met another issue:

 File "/opt/anaconda3/envs/env-37/lib/python3.9/site-packages/Hawkes/model.py", line 849, in set_data
    dl,dInt = preprocess_data_nonpara_cython(Data['T'],self.bin_edge,itv[1])

when I set the baseline as 'plinear' and the kernel as 'nonpara'.

Minimization for Exponential Kernel

Hi, more than an issues this is rather a suggestion.
I'm trying minimize the loglikelihood function for a Hawkes process with kernel that is the sum of P exponentials. For work reasons I wrote the loss function by myself and, for each point (\mu, \alpha1, \alpha2, \beta1, \beta2) both the function value and its gradient coincide with yours. However, by using normal strategies from scipy.minimize, most of the time I find point different from the one of your estimator, especially when there are several order of magnitude of difference between the betas.
So I wanted to ask whether you use multiple guess as a starting point or what, thank you for the attention