Regression

Learn.jl provides SVR, linear regression, decision tree regression, and random forest regression. The functions used are all the same, just the parameters change.

Functions

fit!{T<:AbstractFloat}(reg::Regressor, X::Matrix{T}, y::Vector{T})

Fit the regressor reg to the inputs X and the targets y.

Parameters:

• reg – The regressor object carrying parameters. This parameter will be modified by the function. After running fit! reg will contain all information required to make predictions.
• X – X assumes rows for observations and columns as features.
• y – The targets.

predict{T<:AbstractFloat}(reg::Regressor, X::Matrix{T})

Predict values for X using the fitted estimator reg.

Parameters:

• reg – The regressor after fitting with fit!.
• X – Input data with observations in rows and features in columns.

function score{T<:AbstractFloat}(estimator::Regressor, X::Matrix{T}, y_true::Vector; scoring::Union{Function, Void}=r2_score)

Return the score for a set of input observations X and targets y with scoring function scoring.

Parameters:

• estimator – The estimator trained by fit.
• X – The input values with observations as rows and features as columns.
• y_true – The true targets to compare predictions with.
• scoring – Optional scoring function. By default the R2 score is used for regression.

SVR

Implementation of the eps-SVR algorithm. This code uses LIBSVM.jl, which in turn wraps LIBSVM.

SVR(;kernel="rbf", degree=3, gamma=nothing, coef0=0.0, C=1.0, nu=0.5, p=0.1, cache_size=100.0, eps=0.001, shrinking=true, probability_estimates=false, weights=nothing, verbose=false)

For detailed information about the parameters refer to the documentation of LIBSVM.jl and LIBSVM .

reg = SVR(C=100.0)
fit!(reg, X, y)
y_pred = predict(reg, X)
r2_score(y, y_pred)
assert(r2_score(y, y_pred) == score(reg, X, y))

Linear regression

Linear regression, using GLM.jl.

LinearRegression(;fit_intercept::Bool=true, normalize::Bool=false)

reg = LinearRegression()
fit!(reg, X, y)
y_pred = predict(reg, X)
assert(r2_score(y, y_pred) == score(reg, X, y))

Decision tree regression

Decision tree regression, using DecisionTree.jl. Currently, it seems that the algorithm breaks if the input data has only one feature. The issue is reported.

DecisionTreeRegressor()

reg = DecisionTreeRegressor()
fit!(reg, X, y)
y_pred = predict(reg, X)
assert(r2_score(y, y_pred) == score(reg, X, y))

Random forest regression

Random forest regression using DecisionTree.jl. Currently, it seems that the algorithm breaks if the input data has only one feature. The issue is reported.

RandomForestRegressor(;nsubfeatures::Integer=2, ntrees::Integer=5)

reg = RandomForestRegressor()
fit!(reg, X, y)
y_pred = predict(reg, X)
assert(r2_score(y, y_pred) == score(reg, X, y))