MultitargetKNNClassifier

MultitargetKNNClassifier

A model type for constructing a multitarget K-nearest neighbor classifier, based on NearestNeighborModels.jl, and implementing the MLJ model interface.

From MLJ, the type can be imported using

MultitargetKNNClassifier = @load MultitargetKNNClassifier pkg=NearestNeighborModels

Do model = MultitargetKNNClassifier() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in MultitargetKNNClassifier(K=...).

Multi-target K-Nearest Neighbors Classifier (MultitargetKNNClassifier) is a variation of KNNClassifier that assumes the target variable is vector-valued with Multiclass or OrderedFactor components. (Target data must be presented as a table, however.)

Training data

In MLJ or MLJBase, bind an instance model to data with

mach = machine(model, X, y)

OR

mach = machine(model, X, y, w)

Here:

  • X is any table of input features (eg, a DataFrame) whose columns are of scitype Continuous; check column scitypes with schema(X).
  • yis the target, which can be any table of responses whose element scitype is either<:Finite(<:Multiclassor<:OrderedFactorwill do); check the columns scitypes withschema(y). Each column ofy` is assumed to belong to a common categorical pool.
  • w is the observation weights which can either be nothing(default) or an AbstractVector whose element scitype is Count or Continuous. This is different from weights kernel which is a model hyperparameter, see below.

Train the machine using fit!(mach, rows=...).

Hyper-parameters

  • K::Int=5 : number of neighbors
  • algorithm::Symbol = :kdtree : one of (:kdtree, :brutetree, :balltree)
  • metric::Metric = Euclidean() : any Metric from Distances.jl for the distance between points. For algorithm = :kdtree only metrics which are instances of Union{Distances.Chebyshev, Distances.Cityblock, Distances.Euclidean, Distances.Minkowski, Distances.WeightedCityblock, Distances.WeightedEuclidean, Distances.WeightedMinkowski} are supported.
  • leafsize::Int = algorithm == 10 : determines the number of points at which to stop splitting the tree. This option is ignored and always taken as 0 for algorithm = :brutetree, since brutetree isn't actually a tree.
  • reorder::Bool = true : if true then points which are close in distance are placed close in memory. In this case, a copy of the original data will be made so that the original data is left unmodified. Setting this to true can significantly improve performance of the specified algorithm (except :brutetree). This option is ignored and always taken as false for algorithm = :brutetree.
  • weights::KNNKernel=Uniform() : kernel used in assigning weights to the k-nearest neighbors for each observation. An instance of one of the types in list_kernels(). User-defined weighting functions can be passed by wrapping the function in a UserDefinedKernel kernel (do ?NearestNeighborModels.UserDefinedKernel for more info). If observation weights w are passed during machine construction then the weight assigned to each neighbor vote is the product of the kernel generated weight for that neighbor and the corresponding observation weight.
  • output_type::Type{<:MultiUnivariateFinite}=DictTable : One of (ColumnTable, DictTable). The type of table type to use for predictions. Setting to ColumnTable might improve performance for narrow tables while setting to DictTable improves performance for wide tables.

Operations

  • predict(mach, Xnew): Return predictions of the target given features Xnew, which should have same scitype as X above. Predictions are either a ColumnTable or DictTable of UnivariateFiniteVector columns depending on the value set for the output_type parameter discussed above. The probabilistic predictions are uncalibrated.
  • predict_mode(mach, Xnew): Return the modes of each column of the table of probabilistic predictions returned above.

Fitted parameters

The fields of fitted_params(mach) are:

  • tree: An instance of either KDTree, BruteTree or BallTree depending on the value of the algorithm hyperparameter (See hyper-parameters section above). These are data structures that stores the training data with the view of making quicker nearest neighbor searches on test data points.

Examples

using MLJ, StableRNGs

## set rng for reproducibility
rng = StableRNG(10)

## Dataset generation
n, p = 10, 3
X = table(randn(rng, n, p)) ## feature table
fruit, color = categorical(["apple", "orange"]), categorical(["blue", "green"])
y = [(fruit = rand(rng, fruit), color = rand(rng, color)) for _ in 1:n] ## target_table
## Each column in y has a common categorical pool as expected
selectcols(y, :fruit) ## categorical array
selectcols(y, :color) ## categorical array

## Load MultitargetKNNClassifier
MultitargetKNNClassifier = @load MultitargetKNNClassifier pkg=NearestNeighborModels

## view possible kernels
NearestNeighborModels.list_kernels()

## MultitargetKNNClassifier instantiation
model = MultitargetKNNClassifier(K=3, weights = NearestNeighborModels.Inverse())

## wrap model and required data in an MLJ machine and fit
mach = machine(model, X, y) |> fit!

## predict
y_hat = predict(mach, X)
labels = predict_mode(mach, X)

See also KNNClassifier