Merge objects together.

[raphaelsty]

Hi,

This discussion to ask if it would be relevant to be able to merge two river objects.

I see a use case in the context of running several independent processes. A user might want to merge the objects at regular intervals. I'm not sure if it's relevant or if it fits with the river philosophy. It is obvious to implement for the mean statistic but not sure if it is simple for other methods like linear models.

from river import stats

a = stats.Mean()
a.update(1)

b = stats.Mean()
b.update(4)

c = a.join(b)
c.get()
# 2.5

Maybe the question has already been asked :)

[MaxHalford]

Thanks to our friend from Brazil @smastelini we can do this:

from river import stats

a = stats.Mean()
a.update(1)

b = stats.Mean()
b.update(4)

c = a + b
c.get()

:)

[MaxHalford]

We should document this feature really.

[smastelini]

I totally agree! It is on my to-do list to create a guide for this kind of operation and also another one for warm_up_mode and pure_inference_mode.

About merging objects, I think with some reading about federated learning, we could make stuff like:

lm_a = linear_model.LinearRegression()
lm_b = linear_model.LinearRegression()
# lm_a and lm_b are trained with different samples of a stream
# (maybe in different nodes in a distributed computing setting)

and then

lm_c = lm_a + lm_b

And we have a new model whose weights are a combination of the ones in models A and B.

Sounds like the future, huh?

[raphaelsty]

Amazing! Of course it's the future! 🥳