Multivariate Poisson as marked #51

jucheval · 2025-03-14T10:50:36Z

The goal is to merge the two concrete types MarkedPoissonProcess and MultivariatePoissonProcess as a new concrete type called PoissonProcess (hence it more or less corresponds to the old AbstractPoissonProcess)

At this stage, the two implementations coexist. In particular, there is the test file multivariate_as_marked.jl that checks that the old implementation of multivariate Poisson is equivalent to the new one.

To close this PR, it remains to:

remove the files src/poisson/abstract_poisson_process.jl and test/multivariate_as_marked.jl,
remove and the folders poisson/multivariate and poisson/marked,
fix the TODOs (related to aliases which would be in conflict in the current version).

codecov · 2025-03-14T10:52:03Z

Codecov Report

All modified and coverable lines are covered by tests ✅

📢 Thoughts on this report? Let us know!

jucheval · 2025-03-17T14:45:20Z

The code coverage is worse than before because of the twin implementation of MultivariatePoissonProcess

gdalle

Thanks, and sorry for the review! It needs a few improvements but you can probably remove duplicate code now

src/poisson/fit.jl

src/poisson/multivariate/fit.jl

gdalle · 2025-03-28T11:26:58Z

src/poisson/poisson_process.jl

+struct PoissonProcess{R<:Real,D} <: AbstractPointProcess
+    λ::R
+    mark_dist::D
+    PoissonProcess{R,D}(λ::R, mark_dist::D) where {R,D} = new{R,D}(λ, mark_dist)


This constructor is not needed, it is defined by default

Once again, I adapted it from Distributions.jl. If I understand correctly it is needed so that the (Any, Any) constructor is not defined by default. If I remove it, then julia gives a warning of method redefinition at the first constructor line 31.

I replaced this old inner constructor by the one that checks if the intensity is positive.

gdalle · 2025-03-28T11:28:10Z

src/poisson/poisson_process.jl

+function PoissonProcess(λ::Integer, mark_dist; check_args::Bool=true)
+    return PoissonProcess(float(λ), mark_dist; check_args=check_args)
+end


Is this necessary?

I adapted what is done for Exponential in Distributions.jl

The design of Distributions.jl leaves plenty to be desired and is fairly old, so it shouldn't be a universal reference. Let's avoid automatic conversions for the sake of simplicity

gdalle · 2025-03-28T11:28:56Z

src/poisson/poisson_process.jl

+# TODO: Replace PoissonProcess{R,Categorical{R,Vector{R}}} by
+# MultivariatePoissonProcess{R}
+# when everything else is OK
+function intensity_vector(pp::PoissonProcess{R,Categorical{R,Vector{R}}}) where {R}


Is this generalizable to other kinds of processes?

Also it needs to be documented since it is public

Do you mean that it could be a function defined for some kind of "AbstractMultivariatePointProcess" ?

Or at least document it for the Poisson process

I added a docstring

gdalle · 2025-03-28T11:29:34Z

src/poisson/poisson_process.jl

+    return log(ground_intensity(pp)) + logdensityof(mark_distribution(pp), m)
+end
+
+### Conversions


Are these necessary?

Once again, I adapted this from Distributions.jl. I would say that it is necessary as soon as you have the automatic conversion from integer to float (one of your other comments).

Let's remove them and stay simple. If a user tries to create a Poisson process with an integer intensity and it errors, that's on them.

src/poisson/prior.jl

jucheval · 2025-05-07T18:16:22Z

src/poisson/poisson_process.jl

+
+`MultivariatePoissonProcess{R}` is simply a type alias for `PoissonProcess{R,Categorical{R,Vector{R}}}`.
+"""
+const MultivariatePoissonProcess{R<:Real} = PoissonProcess{R,Categorical{R,Vector{R}}}


Just realized that the automatic conversion made by Distributions.jl (related to other discussions above) makes, for instance, that PoissonProcess(Int[1,1]) isa MultivariatePoissonProcess{Int} is false.
If this "problem" happens only in the case of integers intensities, we could just don't care, but it may also arises to other Real's, no ?

Distributions.jl has a lot of issues related to types. For my package HiddenMarkovModels.jl, I actually rolled out my own version of Categorical, only to avoid them. There was a project called MeasureTheory.jl to rebuild Distributions.jl on a saner basis, but I think it's dead at the moment.

Either we decide that type issues are not our fault if they come from Distributions.jl, or we fix them ourselves. I'm not sure which is the best approach.

A possibility would be to write

Suggested change

const MultivariatePoissonProcess{R<:Real} = PoissonProcess{R,Categorical{R,Vector{R}}}

const MultivariatePoissonProcess{R<:Real} = PoissonProcess{R,Categorical{T,Vector{T}}} where T

The suggestion above is not satisfactory but it makes MultivariatePoissonProcess{R} a UnionAll and not a DataType.

From what I understand from Categorical in Distributions.jl, the type parameter is actually associated with the type of the probabilities in the probability vector. So, the type must be an AbstractFloat, since it is a number between 0 and 1. Except, for example, in the case Categorical([1, 0]), this returns a Categorical{Int64, Vector{Int64}}.

If the intensities are integers, the type for Categorical will still be an AbstratFloat, since it will be the type of the resulting probabilities. What seems to work is

const MultivariatePoissonProcess{R<:Real} = PoissonProcess{R,Categorical{F,Vector{F}}} where {F<:AbstractFloat}

Here is an other suggestion :

const MultivariatePoissonProcess{R<:Real, F} = PoissonProcess{R,Categorical{F,Vector{F}}}

Hence, this is a DataType.

The drawback is that you need to provide e.g. MultivariatePoissonProcess{Float32, Float32} instead of MultivariatePoissonProcess{Float32} to the fit function.

Don't we need to define a separate fit method anyway (or, at least, suffstats)? From what I understand, MultivariatePoissonProcess is for non-marked events, so when we call fit with some vector of histories, we cannot use the marks in the histories, we need build the marks so that each event time has a mark corresponding to which history it came from.

If that is the case, couldn't we just define the method fit(pp::Type{<:MultivariatePoissonProcess{R}}, hists::Vector{<:History}) where {R<:Real}, instead of using MultivariatePoissonPoisson{R,F}? Or have both, but the first just dispatches to the second one with F = float(R)?

For now, fit is implemented by providing the marks, even for multivariate processes. That is one of the reasons why I tried to unify multivariate and marked frameworks with this PR.

FYI, the "main" fit method is :

function StatsAPI.fit( ::Type{PoissonProcess{R,D}}, ss::PoissonProcessStats{R1,R2} ) where {R<:Real,D,R1<:Real,R2<:Real} λ = convert(R, ss.nb_events / ss.duration) mark_dist = fit(D, ss.marks, ss.weights) return PoissonProcess(λ, mark_dist) end

Leveraging on your idea, I would suggest to add this method

function StatsAPI.fit( ::Type{MultivariatePoissonPoisson{R}}, ss::PoissonProcessStats{R1,R2} ) where {R<:Real,R1<:Real,R2<:Real} λ = convert(R, ss.nb_events / ss.duration) mark_dist = fit(Categorical, ss.marks, ss.weights) return PoissonProcess(λ, mark_dist) end

Here, we let Distributions.jl chose the concrete Categorical type.

For now, fit is implemented by providing the marks, even for multivariate processes. That is one of the reasons why I tried to unify multivariate and marked frameworks with this PR.

Makes sense. We can make a convenience function afterwards to build a mulitvariate history, if we see fit.

My take is that Distributions.jl will stop you from propagating any fancy types either way, so introducing a distinction between the type you want (R) and the type you get after conversion (F) is only giving you false hopes

I was playing around with the fit method for Categorical and it does not let you choose the type anyway. fit(Categorical, [1,2,2,1]) and fit(Categorical{BigFloat, Vector{BigFloat}}, [1,2,2,1]) both return a Categorical{Float64, Vector{Float64}}, so this choice of type parameter is an illusion.

So we could do

function StatsAPI.fit( ::Type{MultivariatePoissonPoisson{R}}, ss::PoissonProcessStats{R1,R2} ) where {R<:Real,R1<:Real,R2<:Real} return fit(MultivariatePoisson{R, Float64}, ss) end

or just force F to be a Float64 in the constructor, which only affects the precision of the probabilities of each mark, and this problem with fit goes away.

If propagation of fancy types is impossible due to Distributions.jl as @gdalle suggested, I would say that the better and simpler is to force Float64 in the constructor.

I can make a commit that does that if you want.

I think this is the better solution too. Maybe just leave a comment somewhere saying that this is the reason we forced the type, just to let whoever may be working on the code afterwards know about this.

jucheval · 2025-07-17T11:30:21Z

@gdalle I think that every conversation can be resolved except the last one concerning MultivariatePoissonProcess for which I added a suggestion.

jucheval · 2025-11-14T13:07:00Z

I thought that other parts of the code needed modification but it modifying the alias suffices.
FYI, autodiff seems to work still: line 40 of test/multivariate_poisson_process.jl, i.e.

gf = ForwardDiff.gradient(f1, 3 * ones(10))

does not throw an error.

@JoseKling : do you want to handle the merging ?

JoseKling · 2025-11-14T13:27:50Z

great.

And I can handle it, no problem

JoseKling · 2025-11-14T15:09:45Z

Merged manually via the command line in commit 8dc41c3.

jucheval added 18 commits February 13, 2025 12:41

remove M

62a96a9

add mark_type(pp)

795ae13

remove mark_type

aea381d

fix change of mark type

f9d21b3

init

8c2053c

add PoissonProcess and Dirac

246250a

specialize rand to PoissonProcess instead of AbstractPoissonProcess

be75db1

add file to test equivalence between multivariate and marked

57534ca

add includes

a44a709

typo corrected

c3afbbc

replace MarkedPoissonProcess by PoissonProcess

477a407

add intensity_vector

7920608

implement fit_map for MultivariatePoissonProcess only

d6a4709

update signatures

e47c99c

add check_args to PoissonProcess constructors

e4194cf

formatting

ddbe6c3

Merge branch 'main' into multivariate-as-marked

07da8ba

add mark_distribution method for PoissonProcess

62253b4

add PoissonProcess(Prior) to API

8133173

jucheval requested a review from gdalle March 17, 2025 14:45

gdalle requested changes Mar 28, 2025

View reviewed changes

jucheval added 7 commits March 28, 2025 16:15

remove duplicate code and introduce MultivariatePoissonProcess alias

f266ff9

avoids allocation in sum

ea026b9

fix Base.show methods

78b2da1

fix exports

1a06292

fix multivariate Poisson with null intensities

efec261

fix typos

d0a1c95

take a marked example different from the multivariate one

d5f52be

jucheval added 6 commits March 28, 2025 17:00

add docstring to uni- and multi- variatePoisson

e25cf8b

add Univariate to API

e5b118f

fix MethodError issue

e472e20

consistence of formatting

2c607f4

removes handling of PoissonProcess with intensity of type Int

ec801f0

add docstring to intensity_vector

cd7bbc3

jucheval force-pushed the multivariate-as-marked branch from b7b5b3c to cd7bbc3 Compare May 6, 2025 13:27

jucheval commented May 7, 2025

View reviewed changes

add unit tests to improve coverage

84f867d

jucheval mentioned this pull request Nov 11, 2025

HomogeneousPoissonProcess type #63

Closed

force Categorical{Float64} for MultivariatePoissonProcess

78e1ccc

JoseKling closed this Nov 14, 2025

	const MultivariatePoissonProcess{R<:Real} = PoissonProcess{R,Categorical{R,Vector{R}}}
	const MultivariatePoissonProcess{R<:Real} = PoissonProcess{R,Categorical{T,Vector{T}}} where T

Multivariate Poisson as marked #51

Multivariate Poisson as marked #51

Uh oh!

Conversation

jucheval commented Mar 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Mar 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

jucheval commented Mar 17, 2025

Uh oh!

gdalle left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jucheval May 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jucheval May 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jucheval commented Jul 17, 2025

Uh oh!

jucheval commented Nov 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jucheval commented Mar 14, 2025 •

edited

Loading

codecov bot commented Mar 14, 2025 •

edited

Loading

jucheval May 6, 2025 •

edited

Loading

jucheval May 7, 2025 •

edited

Loading

jucheval commented Nov 14, 2025 •

edited

Loading