Align on fading/smoothing/forgetting terminology

[MaxHalford]

We're moving more and more towards models and algorithms that account for drift. This implies "forgetting" the past and focusing on recent data. Often the mechanism through which this happens is a smoothing factor, typically comprised in [0, 1]. We can see it in EWMean, LossyCount, TextClust, and at some point BayesianLinearRegression will have it too.

This is great news, and it shows we're moving in the right direction of helping people deal with drift.

The one thing that irks me is that we're using different terminology in different places to refer to the same concept. EWMean calls this alpha, LossyCount calls this forgetting, TextClust calls this fading. There are probably other cases but I can't remember them off the top of my head.

I would like to suggest we use a single name for this kind of parameter. I've suggested a few options here, but feel free to suggest something else and to comment -- ping @online-ml/devs.

Once we agree I'll add this to our roadmap.

[raphaelsty]

What about shift or memory or drift ? 🥸

Disadvantage is that shift could indicate an offset (as with the pandas library) and memory could rather make us think of RAM memory.

Sorry but not sorry to add noise

[MaxHalford]

• shift yes it sounds like an offset
• memory that sounds like the window_size parameter. Maybe memory is a better term than window_size
• drift sounds like concept drift, which is the problem the parameter we're discussing is suppose to tackle. I would rather use a different name for the parameter

[Dennis1989]

I don't know of it is possible to really agree upon one term for that. We should keep in mind that there are different approaches when it comes to forgetting data. Different window modes can be for example used (landmark, sliding, damped). Using one fixed term may actually add confusion.

[smastelini]

I agree partially with you. The specific examples @MaxHalford presented specifically concern exponentially fading weights. I think for those we can agree in a single term. But other cases should be handled separately.

[MaxHalford]

Yes I'm talking about fading windows, not regular windows.

[thorbjornwolf]

That sounds like a very good idea!

So, uh, let me also add to the noise from too many options 😄
I've always liked halflife for exponential weights, but maybe that's because of the game.
pandas's ewm has all of com (center of mass), span, halflife and alpha, which is amazing but seems like too much to put in every relevant signature in river.

One wish: I hope the chosen name and its value correspond intuitively. For instance,

• an increase in forgetting should reduce the halflife
• an increase in fading_factor should reduce the halflife
• an increase in smoothing should increase the halflife
• an increase in alpha should... uhhh...

[MaxHalford]

Thanks that's helpful, especially the Halflife reference haha.

[MaxHalford]

Ok let's go with fading_factor! It makes sense the more I think about it. The factor part also makes it clear it's a parameter between 0 and 1. I'll make a pull request to make the necessary changes.

Thanks all!

[Dennis1989]

Sounds good. No work for me adjusting textclust ;)

[MaxHalford]

Haha right. About textClust, it's still on my todo to look into refactoring the code, I just haven't prioritized it yet :)