Note: This glossary includes some detail intended mainly for MLJ developers.
Parameters on which some learning algorithm depends, specified before the algorithm is applied, and where learning is interpreted in the broadest sense. For example, PCA feature reduction is a "preprocessing" transformation "learning" a projection from training data, governed by a dimension hyperparameter. Hyperparameters in our sense may specify configuration (eg, number of parallel processes) even when this does not affect the end-product of learning. (But we exclude verbosity level.)
Object collecting together hyperpameters of a single algorithm. Models are classified either as supervised or unsupervised models (eg, "transformers"), with corresponding subtypes
Supervised <: Model and
Unsupervised <: Model.
Also known as "learned" or "fitted" parameters, these are "weights", "coefficients", or similar parameters learned by an algorithm, after adopting the prescribed hyper-parameters. For example, decision trees of a random forest, the coefficients and intercept of a linear model, or the projection matrices of a PCA dimension-reduction algorithm.
Data-manipulating operations (methods) using some
fitresult. For supervised learners, the
predict_mode methods; for transformers, the
inverse_transform method. An operation may also refer to an ordinary data-manipulating method that does not depend on a fit-result (e.g., a broadcasted logarithm) which is then called static operation for clarity. An operation that is not static is dynamic.
An object consisting of:
(1) A model
(2) A fit-result (undefined until training)
(3) Training arguments (one for each data argument of the model's associated
fit method). A training argument is data used for training (subsampled by specifying
fit!) but also in evaluation (subsampled by specifying
predict_mean, etc). Generally, there are two training arguments for supervised models, and just one for unsupervised models. Each argument is either a
Source node, wrapping concrete data supplied to the
machine constructor, or a
Node, in the case of a learning network (see below). Both kinds of nodes can be called with an optional
rows=... keyword argument to (lazily) return concrete data.
In addition, machines store "report" metadata, for recording algorithm-specific statistics of training (eg, an internal estimate of generalization error, feature importances); and they cache information allowing the fit-result to be updated without repeating unnecessary information.
Machines are trained by calls to a
fit! method which may be passed an optional argument specifying the rows of data to be used in training.
For more, see the Machines section.
Note: Multiple machines in a learning network may share the same model, and multiple learning nodes may share the same machine.
A container for training data and point of entry for new data in a learning network (see below).
Essentially a machine (whose arguments are possibly other nodes) wrapped in an associated operation (e.g.,
inverse_transform). It consists primarily of:
- An operation, static or dynamic.
- A machine, or
nothingif the operation is static.
- Upstream connections to other nodes, specified by a list of arguments (one for each argument of the operation). These are the arguments on which the operation "acts" when the node
Nis called, as in
A directed acyclic graph implicit in the connections of a collection of source(s) and nodes.
Any model with one or more other models as hyper-parameters.
Any wrapper, or any learning network, "exported" as a model (see Composing Models).