Don't use "model" to describe hyperparameter struct.

[ablaom]

Inconsistent with use in Statistics has raised objections. See e.g. https://discourse.julialang.org/t/ann-learnapi-jl-proposal-for-a-basement-level-machine-learning-api/

The question is what to call a struct whose:

• name represents a machine learning algorithm (such as decision tree classifier), and
• whose fields represent the hyperparameters

Suggestions include hyper options, strategy, algorithm

[KronosTheLate]

Is Learner fitting at all?

[davidbp]

I think that ModelHyperparams vs Model makes a separation of concerns:

• ModelHyperparms refers to model hyperparams
• Model refer to model + learned params.

[ablaom]

Is Learner fitting at all?

I like "learner" and I've seen "learner" used in ML papers to refer to the unfitted object. But MLR/MLR3 uses it for the class that also stores learned parameters.

But maybe that's okay. At least we're not likely to get into as much trouble over "learner" as "model".

[ablaom]

After further thought (and I got as far as writing a new PR) I am less fond of "learner". The problem is that plenty of machine learning algorithms don't really learn (don't generalize to new data).

We've had a fair bit of discussion about this here and in other threads. I think consensus is impossible and have decided to go with "algorithm".

[KronosTheLate]

We've had a fair bit of discussion about this here and in other threads. I think consensus is impossible and have decided to go with "algorithm".

Fair enough. Consensus indeed seems impossible, and debating naming is not well spent time.

[ablaom]

Re-opening.

What are people's preferences? "learner", "strategy", or "algorithm"? Please vote here: https://discourse.julialang.org/t/ann-learnapi-jl-proposal-for-a-basement-level-machine-learning-api/93048/31?u=ablaom

[davidbp]

From the provided options:

• ModelStragery refers to model hyperparameres or "stragegy to use during learning"
• Model refers to model hyperparams + learned params.
  or
• ModelAlgorithm refers to model hyperparameres or "stragegy to use during learning"
• Model refers to model hyperparams + learned params.

I understand some people might want type names that are very simple but Strategy or Algorithm seem just too generic.

For example: the package NearestNeighbors defines a bunch of "tree models" such as BruteTree, KDTree, BallTree, DataFreeTree all of them subtypes of the abstractNNTree. It could have been just Tree but the NN tells you what the subtypes might be doing "nearest neighbors".

In a similar way adding Model or Learner as a prefix to whatever name you decide to use for the struct will tell users how is that struct expected to be used (for example LearnerStragety, LearnerHyperParams, LearnerSchema or ModelStragety, ModelHyperParams, ModelSchema) and if that is too long one could always use MStrategy or LStrategy to refer to ModelStrategy or LearnerStrategy respectively.

[ablaom]

After the poll at https://discourse.julialang.org/t/ann-learnapi-jl-proposal-for-a-basement-level-machine-learning-api/93048/61, I will change "algorithm" to "stategy".

[ablaom]

With the greatest respect for the alternatives presented, I'm sticking with "algorithm". After many re-workings of the docs, this creates the least cognitive dissonance for me personally, by a considerable amount, despite not being my original first choice.