HSGP as component #1246

wd60622 · 2024-11-30T11:45:20Z

Description

Adding HSGP component which acts a bit more like Prior class than MMM
components but with some plot_curve logic as well.

Bringing back the PR that @bwengals started for better priors for the 1d HSGP

import numpy as np
import pandas as pd

import matplotlib.pyplot as plt

from pymc_marketing.mmm import HSGP

n = 52
X = np.arange(n)

hsgp = HSGP.parameterize_from_data(X=X, dims="time")

dates = pd.date_range("2024-01-01", periods=n, freq="W-MON")
coords = {"time": dates}
prior = hsgp.sample_prior(coords=coords)
curve = prior["f"]
fig, axes = hsgp.plot_curve(curve)
ax = axes[0]
ax.set(xlabel="Time", ylabel="f", title="Out of the box HSGP prior")
plt.show()

EDIT: I've implemented this with the HSGP.parameterize_from_data class method:

We might want to refactor this to pull out the defaults instead of having them
as property of the class. Maybe like:

eta: Prior = get_recommended_eta(X)
ell: Prior = get_recommended_ell(X)
hsgp = HSGP(dims="time", eta=eta, ell=ell)

Related Issue

Closes Pull out TVP as component #751
Related to Better default priors for TVP & TVP component #829

Checklist

Checked that the pre-commit linting/style checks pass
Included tests that prove the fix is effective or that the new feature works
Added necessary documentation (docstrings and/or example notebooks)
If you are a pro: each commit corresponds to a relevant logical change

Modules affected

MMM
CLV

Type of change

📚 Documentation preview 📚: https://pymc-marketing--1246.org.readthedocs.build/en/1246/

codecov · 2024-11-30T12:08:15Z

Codecov Report

Attention: Patch coverage is 93.43434% with 26 lines in your changes missing coverage. Please review.

Project coverage is 95.02%. Comparing base (f4d5030) to head (142b107).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
pymc_marketing/mmm/hsgp.py	91.62%	17 Missing ⚠️
pymc_marketing/plot.py	96.40%	6 Missing ⚠️
pymc_marketing/mmm/plot.py	0.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1246      +/-   ##
==========================================
- Coverage   95.27%   95.02%   -0.26%     
==========================================
  Files          40       42       +2     
  Lines        4234     4439     +205     
==========================================
+ Hits         4034     4218     +184     
- Misses        200      221      +21

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

wd60622 · 2024-11-30T16:57:57Z

Checkout the documentation section here:
https://pymc-marketing--1246.org.readthedocs.build/en/1246/api/generated/pymc_marketing.hsgp_kwargs.HSGP.html#pymc_marketing.hsgp_kwargs.HSGP

Thoughts on using the create_variable over apply as GP is thought to be a distribution over functions wheras apply is a bit more for transformation of data.

One aspect that is tough is that the new parametrization is different than the HSGPKwargs class and conflicts a bit. This would be a breaking change and require a version increase for MMM class. However, this might be another argument for parameterizing the HSGP with distributions which might be more flexible in the long run. I think there are still some conflicts with this route due to the nature of this.

Would people want to do something like this?

from pymc_marketing.prior import Prior

eta = Prior("HalfNormal", sigma=Prior("HalfNormal"), dims="channel")
ell = Prior("HalfNormal", dims="channel")
hsgp = HSGP(dims=("time", "channel"), eta=eta, ell=ell)

TODO:

Add to the components notebook
Write test suite for the HSGP component
Incorporate into the MMM directly after determining relationship between HSGP and HSGPKwargs currently used.

@cetagostini I was mistaken when I talked out it being hierachical, it is not.
The coefs are draw independently but the basis are all the same which confused me.

pymc-marketing/pymc_marketing/hsgp_kwargs.py

Lines 537 to 543 in f8975a9

    
           hsgp_coefs = Prior( 
        
               "Normal", 
        
               mu=0, 
        
               sigma=sqrt_psd, 
        
               dims=hsgp_dims, 
        
               centered=self.centered, 
        
           ).create_variable(f"{name}_hsgp_coefs")

wd60622 · 2024-11-30T17:19:50Z

Question for @bwengals :
Is there a benefit of centered vs non-centered since sqrt_psd will be a vector and not a random vector of size (m, )? hsgp_coefs will potentially have more dims (*other_dims, m)

juanitorduz · 2024-11-30T18:48:21Z

This looks very nice! In the long run, this seems like a nicer approach. As we have not reached a v1.0.0 version I think it's fine that the API changes from time to time

juanitorduz

This looks great @wd60622 ! I really like the proposal! A couple of questions and comments:

Do we need more tests for the hsgp_kwargs? What is missing in this PR?
It would be great to update the GP MMM notebooks (maybe different PR) to showcase this component point of view.
Can we also do HSGPPeriodic? I think modeling the media time-varying coefficients as a periodic GP is something we have been asked for.

pymc_marketing/hsgp_kwargs.py

juanitorduz · 2024-11-30T18:51:00Z

pymc_marketing/hsgp_kwargs.py

+
+
+def pc_prior_1d(alpha: float = 0.1, lower: float = 1.0) -> Prior:
+    R"""Create a one-dimensional PC prior for GP lengthscale.


Can we describe what "PC" stands for?

Do we have any references we can add on this?

Penalized complexity. Here's the reference: https://arxiv.org/abs/1503.00256

It'd be nice to have the PC prior available, but IMO a better default prior for PyMC marketing would be InverseGamma, and then have the user give a lower and upper bound, as described here: https://betanalpha.github.io/assets/case_studies/gaussian_processes.html

I think ideally:

user specifies a lower bound and prior mass -> PC prior

user specifies a lower, upper bound and mass -> InverseGamma

pymc_marketing/hsgp_kwargs.py

juanitorduz · 2024-11-30T18:54:20Z

pymc_marketing/hsgp_kwargs.py

+    )
+
+
+def approx_hsgp_hyperparams(


I thought this function was already available in pymc (gp.hsgp_approx.py)

Doesn't seem to be in pymc 5.13.0: https://github.com/pymc-devs/pymc/blob/v5.13.0/pymc/gp/hsgp_approx.py#L603

@juanitorduz

wd60622 · 2024-11-30T20:33:53Z

This looks great @wd60622 ! I really like the proposal! A couple of questions and comments:

Do we need more tests for the hsgp_kwargs? What is missing in this PR?

To do above

It would be great to update the GP MMM notebooks (maybe different PR) to showcase this component point of view.

Part of todo for the components notebook. Need to think about how this integrates into MMM and also MMM notebook

Can we also do HSGPPeriodic? I think modeling the media time-varying coefficients as a periodic GP is something we have been asked for.

Do you know if the heuristic still works for that one as well?

review-notebook-app · 2024-12-02T16:10:33Z

Check out this pull request on

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

review-notebook-app · 2024-12-02T17:40:49Z

View / edit / reply to this conversation on ReviewNB

cetagostini commented on 2024-12-02T17:40:49Z
----------------------------------------------------------------

Line #9.    media_tvp_config = {

This should change to something like:

python

edia_tvp_config = {

"media_tvp_config": HSGPKwargs(..)

}

pymc_marketing/mmm/tvp.py

cetagostini · 2024-12-02T17:46:32Z

Thoughts on the location of this logic? hsgp_kwargs doesn't seem like a good name and might be good in the mmm module.

Indeed, maybe something related to latent makes more sense to me, or time. If this is a component already, should be under component folders, and the name should be related to what are we trying to model.

We want to model saturation, then we have the saturation component, if we want to model adstock, then we have the adstock component, same thing if we want to model time/latent effects.

cetagostini · 2024-12-02T17:48:06Z

Question for @bwengals : Is there a benefit of centered vs non-centered since sqrt_psd will be a vector and not a random vector of size (m, )? hsgp_coefs will potentially have more dims (*other_dims, m)

I think should be possible, but better to have @bwengals input here 😅

juanitorduz · 2024-12-02T20:28:23Z

hsgp_kwargs

Yes! hsgp_kwargs is a bad name (my fault). You (@wd60622 ) are much better than me at it XD Feel free to push a suggestion

bwengals · 2024-12-02T21:15:58Z

Is there a benefit of centered vs non-centered since sqrt_psd will be a vector and not a random vector of size (m, )? hsgp_coefs will potentially have more dims (*other_dims, m)

Sorry i missed this. Yes sometimes you might want to use the centered parameterization, though non-centered is probably the better default.

juanitorduz · 2024-12-05T19:33:55Z

pymc_marketing/mmm/hsgp.py

+    )
+
+
+def create_m_and_L_recommendations(


Can't we use https://github.com/pymc-devs/pymc/blob/main/pymc/gp/hsgp_approx.py#L97 ?

Can we add some references on where these heuristic come from (just the paper reference, ideally the section)

(I think we can use the same ref as above)

juanitorduz · 2024-12-05T19:34:21Z

pymc_marketing/mmm/hsgp.py

+    ).constrain(lower=lower, upper=upper, mass=mass)
+
+
+def approx_hsgp_hyperparams(


Is it the same as in https://github.com/pymc-devs/pymc/blob/main/pymc/gp/hsgp_approx.py#L97 ?

Not in the min pymc version that we support. How should we handle?

pymc_marketing/mmm/hsgp.py

juanitorduz

I actually think this is great for a first iteration and I think we can improve this over time :)

I just left some comments on the references and some apparent dupplicated code from pymc

wd60622 · 2024-12-05T20:43:17Z

Is there a benefit of centered vs non-centered since sqrt_psd will be a vector and not a random vector of size (m, )? hsgp_coefs will potentially have more dims (*other_dims, m)

Sorry i missed this. Yes sometimes you might want to use the centered parameterization, though non-centered is probably the better default.

Need to double check which I am using as the default. I will set non-centered as default

wd60622 · 2024-12-05T20:43:37Z

What is the status on drop_first? I am getting warnings

juanitorduz · 2024-12-05T20:45:42Z

What is the status on drop_first? I am getting warnings

They will probably go after we support the latest versions (some important fixes were made)

juanitorduz · 2024-12-11T18:38:20Z

@wd60622 what is the status on this one :) ?

wd60622 · 2024-12-11T19:07:09Z

@wd60622 what is the status on this one :) ?

Ready to merge. Will add to example notebooks later

wd60622 · 2024-12-12T02:46:45Z

@wd60622 what is the status on this one :) ?

Ready to merge. Will add to example notebooks later

Later like in another PR

juanitorduz · 2024-12-12T10:48:55Z

Merged ! Cool stuff @wd60622 !

wd60622 added 2 commits November 30, 2024 12:09

allow for non-centered with scalers

ee530d6

bring back the logic from #829

6fb3278

github-actions bot added tests Prior class MMM enhancement New feature or request labels Nov 30, 2024

wd60622 added 3 commits November 30, 2024 14:39

closes #1177

9a7ae6f

import plot_curve

7168348

support for more dimensions

c6e9623

wd60622 linked an issue Nov 30, 2024 that may be closed by this pull request

Move pymc_marketing.mmm.plot module to pymc_marketing.plot module #1177

Closed

wd60622 added 4 commits November 30, 2024 17:42

move the test suite

a182b93

change the import in the example

06b4d15

add back the type hint

7920aee

hopefully display the plots

f14e44a

github-actions bot added maintenance plots labels Nov 30, 2024

correct the docstring

f8975a9

more the tests around

3cf6ce0

wd60622 requested review from juanitorduz and cetagostini November 30, 2024 17:11

juanitorduz reviewed Nov 30, 2024

View reviewed changes

wd60622 added 4 commits December 1, 2024 15:38

change the reference format

c12e431

implement HSGPPeriod based on prior_linearized example

89dd688

Be explicit in the docstring

0bd3c65

Switch to using pytensor for operations

4e7e109

run the notebook with new names

a9f8cc7

github-actions bot added the docs Improvements or additions to documentation label Dec 2, 2024

Merge branch 'main' into bills-hsgp-as-component

09a826f

cetagostini reviewed Dec 2, 2024

View reviewed changes

pymc_marketing/mmm/tvp.py Outdated Show resolved Hide resolved

migrate to mmm.hsgp module

ced4ffe

github-actions bot added the TVPs Related to time varying parameters label Dec 3, 2024

wd60622 and others added 2 commits December 4, 2024 09:28

support for more dimensions in HSGPPeriodic

accea5e

Merge branch 'main' into bills-hsgp-as-component

6b7b6c6

juanitorduz reviewed Dec 5, 2024

View reviewed changes

juanitorduz approved these changes Dec 5, 2024

View reviewed changes

Specific training in description

62ed55f

Merge branch 'main' into bills-hsgp-as-component

142b107

juanitorduz assigned wd60622 Dec 11, 2024

juanitorduz added this to the 0.11.0 milestone Dec 11, 2024

juanitorduz merged commit 45d03ea into main Dec 12, 2024
12 of 13 checks passed

juanitorduz deleted the bills-hsgp-as-component branch December 12, 2024 10:48

wd60622 mentioned this pull request Dec 13, 2024

1271 predictions=true for sample posterior predictive throws error #1272

Merged

13 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

HSGP as component #1246

HSGP as component #1246

wd60622 commented Nov 30, 2024 •

edited

Loading

codecov bot commented Nov 30, 2024 •

edited

Loading

wd60622 commented Nov 30, 2024 •

edited

Loading

wd60622 commented Nov 30, 2024

juanitorduz commented Nov 30, 2024

juanitorduz left a comment

juanitorduz Nov 30, 2024

bwengals Dec 2, 2024

bwengals Dec 2, 2024

juanitorduz Nov 30, 2024

wd60622 Dec 1, 2024

wd60622 Dec 5, 2024

wd60622 commented Nov 30, 2024

review-notebook-app bot commented Dec 2, 2024

review-notebook-app bot commented Dec 2, 2024 •

edited

Loading

cetagostini commented Dec 2, 2024

cetagostini commented Dec 2, 2024

juanitorduz commented Dec 2, 2024

bwengals commented Dec 2, 2024

juanitorduz Dec 5, 2024

juanitorduz Dec 5, 2024

juanitorduz Dec 5, 2024

juanitorduz Dec 5, 2024

wd60622 Dec 5, 2024

juanitorduz left a comment

wd60622 commented Dec 5, 2024

wd60622 commented Dec 5, 2024

juanitorduz commented Dec 5, 2024

juanitorduz commented Dec 11, 2024

wd60622 commented Dec 11, 2024

wd60622 commented Dec 12, 2024

juanitorduz commented Dec 12, 2024



		def pc_prior_1d(alpha: float = 0.1, lower: float = 1.0) -> Prior:
		R"""Create a one-dimensional PC prior for GP lengthscale.

		).constrain(lower=lower, upper=upper, mass=mass)


		def approx_hsgp_hyperparams(

HSGP as component #1246

HSGP as component #1246

Conversation

wd60622 commented Nov 30, 2024 • edited Loading

Description

Related Issue

Checklist

Modules affected

Type of change

codecov bot commented Nov 30, 2024 • edited Loading

Codecov Report

wd60622 commented Nov 30, 2024 • edited Loading

wd60622 commented Nov 30, 2024

juanitorduz commented Nov 30, 2024

juanitorduz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wd60622 commented Nov 30, 2024

review-notebook-app bot commented Dec 2, 2024

review-notebook-app bot commented Dec 2, 2024 • edited Loading

cetagostini commented Dec 2, 2024

cetagostini commented Dec 2, 2024

juanitorduz commented Dec 2, 2024

bwengals commented Dec 2, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

juanitorduz left a comment

Choose a reason for hiding this comment

wd60622 commented Dec 5, 2024

wd60622 commented Dec 5, 2024

juanitorduz commented Dec 5, 2024

juanitorduz commented Dec 11, 2024

wd60622 commented Dec 11, 2024

wd60622 commented Dec 12, 2024

juanitorduz commented Dec 12, 2024

wd60622 commented Nov 30, 2024 •

edited

Loading

codecov bot commented Nov 30, 2024 •

edited

Loading

wd60622 commented Nov 30, 2024 •

edited

Loading

review-notebook-app bot commented Dec 2, 2024 •

edited

Loading