diff --git a/dev/.documenter-siteinfo.json b/dev/.documenter-siteinfo.json index 089b79aa..208330d5 100644 --- a/dev/.documenter-siteinfo.json +++ b/dev/.documenter-siteinfo.json @@ -1 +1 @@ -{"documenter":{"julia_version":"1.6.7","generation_timestamp":"2023-10-05T03:24:51","documenter_version":"1.1.0"}} \ No newline at end of file +{"documenter":{"julia_version":"1.6.7","generation_timestamp":"2023-10-05T04:04:01","documenter_version":"1.1.0"}} \ No newline at end of file diff --git a/dev/api/index.html b/dev/api/index.html index f00dfcb7..11ca3305 100644 --- a/dev/api/index.html +++ b/dev/api/index.html @@ -12,9 +12,9 @@ print_every_n=9999, verbosity=1, return_logger=false, - device="cpu")

Main training function. Performs model fitting given configuration params, dtrain, target_name and other optional kwargs.

Arguments

Keyword arguments

source
fit_evotree(params::EvoTypes{L};
+    device="cpu")

Main training function. Performs model fitting given configuration params, dtrain, target_name and other optional kwargs.

Arguments

Keyword arguments

source
fit_evotree(params::EvoTypes{L};
     x_train::AbstractMatrix, y_train::AbstractVector, w_train=nothing, offset_train=nothing,
     x_eval=nothing, y_eval=nothing, w_eval=nothing, offset_eval=nothing,
     early_stopping_rounds=9999,
     print_every_n=9999,
-    verbosity=1)

Main training function. Performs model fitting given configuration params, x_train, y_train and other optional kwargs.

Arguments

Keyword arguments

source

Predict

MLJModelInterface.predict — Function
predict(model::EvoTree, X::AbstractMatrix; ntree_limit = length(model.trees))

Predictions from an EvoTree model - sums the predictions from all trees composing the model. Use ntree_limit=N to only predict with the first N trees.

source

Features Importance

EvoTrees.importance — Function
importance(model::EvoTree; fnames=model.info[:fnames])

Sorted normalized feature importance based on loss function gain. Feature names associated to the model are stored in model.info[:fnames] as a string Vector and can be updated at any time. Eg: model.info[:fnames] = new_fnames_vec.

source
+ verbosity=1)

Main training function. Performs model fitting given configuration params, x_train, y_train and other optional kwargs.

Arguments

Keyword arguments

source

Predict

MLJModelInterface.predict — Function
predict(model::EvoTree, X::AbstractMatrix; ntree_limit = length(model.trees))

Predictions from an EvoTree model - sums the predictions from all trees composing the model. Use ntree_limit=N to only predict with the first N trees.

source

Features Importance

EvoTrees.importance — Function
importance(model::EvoTree; fnames=model.info[:fnames])

Sorted normalized feature importance based on loss function gain. Feature names associated to the model are stored in model.info[:fnames] as a string Vector and can be updated at any time. Eg: model.info[:fnames] = new_fnames_vec.

source
diff --git a/dev/index.html b/dev/index.html index 50aee9c8..427ae3df 100644 --- a/dev/index.html +++ b/dev/index.html @@ -19,7 +19,7 @@ config = EvoTreeRegressor(rowsample=0.5, rng=123) m2 = fit_evotree(config, df; target_name="y");

However, the following m1 and m2 models won't be because the there's stochasticity involved in the model from rowsample and the random generator in the config isn't reset between the fits:

config = EvoTreeRegressor(rowsample=0.5, rng=123)
 m1 = fit_evotree(config, df; target_name="y");
-m2 = fit_evotree(config, df; target_name="y");

Note that in presence of multiple identical or very highly correlated features, model may not be reproducible if features are permuted since in situation where 2 features provide identical gains, the first one will be selected. Therefore, if the identity relationship doesn't hold on new data, different predictions will be returned from models trained on different features order.

At the moment, there's no reproducibility guarantee on GPU, although this may change in the future.

Missing values

Features

EvoTrees does not handle features having missing values. Proper preprocessing of the data is therefore needed (and a general good practice regardless of the ML model used).

This includes situations where values may be all non-missing, but where the eltype is the form Union{Missing,Float64}. A conversion the types using identity is recommended: 

julia> x = Vector{Union{Missing, Float64}}([1, 2])
+m2 = fit_evotree(config, df; target_name="y");

Note that in presence of multiple identical or very highly correlated features, model may not be reproducible if features are permuted since in situation where 2 features provide identical gains, the first one will be selected. Therefore, if the identity relationship doesn't hold on new data, different predictions will be returned from models trained on different features order.

At the moment, there's no reproducibility guarantee on GPU, although this may change in the future.

Missing values

Features

EvoTrees does not handle features having missing values. Proper preprocessing of the data is therefore needed (and a general good practice regardless of the ML model used).

This includes situations where values may be all non-missing, but where the eltype is Union{Missing,Float64} or Any for example. A conversion using identity is then recommended: 

julia> x = Vector{Union{Missing, Float64}}([1, 2])
 2-element Vector{Union{Missing, Float64}}:
  1.0
  2.0
@@ -27,7 +27,9 @@
 julia> identity.(x)
 2-element Vector{Float64}:
  1.0
- 2.0

For dealing with numerical or ordered categorical features containing missing values, a common approach is to first create an Bool indicator variable capturing the info on whether a value is missing:

transform!(df, :my_feat => ByRow(ismissing) => :my_feat_ismissing)

Then, the missing values can be imputed (replaced by some default values such as mean or median, or using a more sophisticated approach such as predictions from another model):

transform!(df, :my_feat => (x -> coalesce.(x, median(skipmissing(x)))) => :my_feat);

For unordered categorical variables, a recode of the missing into a non missing level is sufficient:

julia> x = categorical(["a", "b", missing])
+ 2.0

For dealing with numerical or ordered categorical features containing missing values, a common approach is to first create an Bool variable capturing the info on whether a value is missing:

using DataFrames
+transform!(df, :my_feat => ByRow(ismissing) => :my_feat_ismissing)

Then, the missing values can be imputed (replaced by some default values such as mean or median, or using a more sophisticated approach such as predictions from another model):

transform!(df, :my_feat => (x -> coalesce.(x, median(skipmissing(x)))) => :my_feat);

For unordered categorical variables, a recode of the missing into a non missing level is sufficient:

using CategoricalArrays
+julia> x = categorical(["a", "b", missing])
 3-element CategoricalArray{Union{Missing, String},1,UInt32}:
  "a"
  "b"
@@ -38,4 +40,4 @@
  "a"
  "b"
  "missing value"

Target

Target variable must have its element type <:Real. Only exception is for EvoTreeClassifier for which CategoricalValue, Integer, String and Char are supported.

Save/Load

EvoTrees.save(m, "data/model.bson")
-m = EvoTrees.load("data/model.bson");
+m = EvoTrees.load("data/model.bson"); diff --git a/dev/models/index.html b/dev/models/index.html index 90e75180..67d71a16 100644 --- a/dev/models/index.html +++ b/dev/models/index.html @@ -11,7 +11,7 @@ model = EvoTreeRegressor(max_depth=5, nbins=32, nrounds=100) X, y = @load_boston mach = machine(model, X, y) |> fit! -preds = predict(mach, X)source

EvoTreeClassifier

EvoTrees.EvoTreeClassifier — Type

EvoTreeClassifier(;kwargs...)

A model type for constructing a EvoTreeClassifier, based on EvoTrees.jl, and implementing both an internal API and the MLJ model interface. EvoTreeClassifier is used to perform multi-class classification, using cross-entropy loss.

Hyper-parameters

  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0. A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.
  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model.
  • gamma::T=0.0: Minimum gain improvement needed to perform a node split. Higher gamma can result in a more robust model. Must be >= 0.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=1.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeClassifier() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeClassifier(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, K] where K is the number of classes:

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeClassifier = @load EvoTreeClassifier pkg=EvoTrees

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

Training data

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

mach = machine(model, X, y)

where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)
  • y: is the target, which can be any AbstractVector whose element scitype is <:Multiclas or <:OrderedFactor; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): return predictions of the target given features Xnew having the same scitype as X above. Predictions are probabilistic.

  • predict_mode(mach, Xnew): returns the mode of each of the prediction above.

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
+preds = predict(mach, X)
source

EvoTreeClassifier

EvoTrees.EvoTreeClassifier — Type

EvoTreeClassifier(;kwargs...)

A model type for constructing a EvoTreeClassifier, based on EvoTrees.jl, and implementing both an internal API and the MLJ model interface. EvoTreeClassifier is used to perform multi-class classification, using cross-entropy loss.

Hyper-parameters

  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0. A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.
  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model.
  • gamma::T=0.0: Minimum gain improvement needed to perform a node split. Higher gamma can result in a more robust model. Must be >= 0.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=1.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeClassifier() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeClassifier(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, K] where K is the number of classes:

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeClassifier = @load EvoTreeClassifier pkg=EvoTrees

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

Training data

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

mach = machine(model, X, y)

where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)
  • y: is the target, which can be any AbstractVector whose element scitype is <:Multiclas or <:OrderedFactor; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): return predictions of the target given features Xnew having the same scitype as X above. Predictions are probabilistic.

  • predict_mode(mach, Xnew): returns the mode of each of the prediction above.

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
 using EvoTrees
 config = EvoTreeClassifier(max_depth=5, nbins=32, nrounds=100)
 nobs, nfeats = 1_000, 5
@@ -24,7 +24,7 @@
 X, y = @load_iris
 mach = machine(model, X, y) |> fit!
 preds = predict(mach, X)
-preds = predict_mode(mach, X)

See also EvoTrees.jl.

source

EvoTreeCount

EvoTrees.EvoTreeCount — Type

EvoTreeCount(;kwargs...)

A model type for constructing a EvoTreeCount, based on EvoTrees.jl, and implementing both an internal API the MLJ model interface. EvoTreeCount is used to perform Poisson probabilistic regression on count target.

Hyper-parameters

  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0. A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.
  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model. Must be >= 0.
  • gamma::T=0.0: Minimum gain imprvement needed to perform a node split. Higher gamma can result in a more robust model.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=1.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • monotone_constraints=Dict{Int, Int}(): Specify monotonic constraints using a dict where the key is the feature index and the value the applicable constraint (-1=decreasing, 0=none, 1=increasing).
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeCount() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeCount(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, 1]:

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeCount = @load EvoTreeCount pkg=EvoTrees

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

Training data

In MLJ or MLJBase, bind an instance model to data with mach = machine(model, X, y) where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)
  • y: is the target, which can be any AbstractVector whose element scitype is <:Count; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): returns a vector of Poisson distributions given features Xnew having the same scitype as X above. Predictions are probabilistic.

Specific metrics can also be predicted using:

  • predict_mean(mach, Xnew)
  • predict_mode(mach, Xnew)
  • predict_median(mach, Xnew)

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
+preds = predict_mode(mach, X)

See also EvoTrees.jl.

source

EvoTreeCount

EvoTrees.EvoTreeCount — Type

EvoTreeCount(;kwargs...)

A model type for constructing a EvoTreeCount, based on EvoTrees.jl, and implementing both an internal API the MLJ model interface. EvoTreeCount is used to perform Poisson probabilistic regression on count target.

Hyper-parameters

  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0. A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.
  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model. Must be >= 0.
  • gamma::T=0.0: Minimum gain imprvement needed to perform a node split. Higher gamma can result in a more robust model.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=1.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • monotone_constraints=Dict{Int, Int}(): Specify monotonic constraints using a dict where the key is the feature index and the value the applicable constraint (-1=decreasing, 0=none, 1=increasing).
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeCount() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeCount(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, 1]:

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeCount = @load EvoTreeCount pkg=EvoTrees

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

Training data

In MLJ or MLJBase, bind an instance model to data with mach = machine(model, X, y) where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)
  • y: is the target, which can be any AbstractVector whose element scitype is <:Count; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): returns a vector of Poisson distributions given features Xnew having the same scitype as X above. Predictions are probabilistic.

Specific metrics can also be predicted using:

  • predict_mean(mach, Xnew)
  • predict_mode(mach, Xnew)
  • predict_median(mach, Xnew)

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
 using EvoTrees
 config = EvoTreeCount(max_depth=5, nbins=32, nrounds=100)
 nobs, nfeats = 1_000, 5
@@ -40,7 +40,7 @@
 preds = predict_mean(mach, X)
 preds = predict_mode(mach, X)
 preds = predict_median(mach, X)
-

See also EvoTrees.jl.

source

EvoTreeMLE

EvoTrees.EvoTreeMLE — Type

EvoTreeMLE(;kwargs...)

A model type for constructing a EvoTreeMLE, based on EvoTrees.jl, and implementing both an internal API the MLJ model interface. EvoTreeMLE performs maximum likelihood estimation. Assumed distribution is specified through loss kwargs. Both Gaussian and Logistic distributions are supported.

Hyper-parameters

loss=:gaussian: Loss to be be minimized during training. One of:

  • :gaussian / :gaussian_mle
  • :logistic / :logistic_mle
  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0.

A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.

  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model.
  • gamma::T=0.0: Minimum gain imprvement needed to perform a node split. Higher gamma can result in a more robust model. Must be >= 0.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=8.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • monotone_constraints=Dict{Int, Int}(): Specify monotonic constraints using a dict where the key is the feature index and the value the applicable constraint (-1=decreasing, 0=none, 1=increasing). !Experimental feature: note that for MLE regression, constraints may not be enforced systematically.
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeMLE() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeMLE(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, nparams] where the second dimensions refer to μ & σ for Normal/Gaussian and μ & s for Logistic.

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeMLE = @load EvoTreeMLE pkg=EvoTrees

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

Training data

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

mach = machine(model, X, y)

where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)

  • y: is the target, which can be any AbstractVector whose element scitype is <:Continuous; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): returns a vector of Gaussian or Logistic distributions (according to provided loss) given features Xnew having the same scitype as X above.

Predictions are probabilistic.

Specific metrics can also be predicted using:

  • predict_mean(mach, Xnew)
  • predict_mode(mach, Xnew)
  • predict_median(mach, Xnew)

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
+

See also EvoTrees.jl.

source

EvoTreeMLE

EvoTrees.EvoTreeMLE — Type

EvoTreeMLE(;kwargs...)

A model type for constructing a EvoTreeMLE, based on EvoTrees.jl, and implementing both an internal API the MLJ model interface. EvoTreeMLE performs maximum likelihood estimation. Assumed distribution is specified through loss kwargs. Both Gaussian and Logistic distributions are supported.

Hyper-parameters

loss=:gaussian: Loss to be be minimized during training. One of:

  • :gaussian / :gaussian_mle
  • :logistic / :logistic_mle
  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0.

A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.

  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model.
  • gamma::T=0.0: Minimum gain imprvement needed to perform a node split. Higher gamma can result in a more robust model. Must be >= 0.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=8.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • monotone_constraints=Dict{Int, Int}(): Specify monotonic constraints using a dict where the key is the feature index and the value the applicable constraint (-1=decreasing, 0=none, 1=increasing). !Experimental feature: note that for MLE regression, constraints may not be enforced systematically.
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeMLE() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeMLE(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, nparams] where the second dimensions refer to μ & σ for Normal/Gaussian and μ & s for Logistic.

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeMLE = @load EvoTreeMLE pkg=EvoTrees

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

Training data

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

mach = machine(model, X, y)

where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)

  • y: is the target, which can be any AbstractVector whose element scitype is <:Continuous; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): returns a vector of Gaussian or Logistic distributions (according to provided loss) given features Xnew having the same scitype as X above.

Predictions are probabilistic.

Specific metrics can also be predicted using:

  • predict_mean(mach, Xnew)
  • predict_mode(mach, Xnew)
  • predict_median(mach, Xnew)

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
 using EvoTrees
 config = EvoTreeMLE(max_depth=5, nbins=32, nrounds=100)
 nobs, nfeats = 1_000, 5
@@ -55,7 +55,7 @@
 preds = predict(mach, X)
 preds = predict_mean(mach, X)
 preds = predict_mode(mach, X)
-preds = predict_median(mach, X)
source

EvoTreeGaussian

EvoTreeGaussian is to be deprecated. Please use EvoTreeMLE with loss = :gaussian_mle.

EvoTrees.EvoTreeGaussian — Type

EvoTreeGaussian(;kwargs...)

A model type for constructing a EvoTreeGaussian, based on EvoTrees.jl, and implementing both an internal API the MLJ model interface. EvoTreeGaussian is used to perform Gaussian probabilistic regression, fitting μ and σ parameters to maximize likelihood.

Hyper-parameters

  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0. A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.
  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model.
  • gamma::T=0.0: Minimum gain imprvement needed to perform a node split. Higher gamma can result in a more robust model. Must be >= 0.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=8.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • monotone_constraints=Dict{Int, Int}(): Specify monotonic constraints using a dict where the key is the feature index and the value the applicable constraint (-1=decreasing, 0=none, 1=increasing). !Experimental feature: note that for Gaussian regression, constraints may not be enforce systematically.
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeGaussian() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeGaussian(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, 2] where the second dimensions refer to μ and σ respectively:

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeGaussian = @load EvoTreeGaussian pkg=EvoTrees

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

Training data

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

mach = machine(model, X, y)

where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)

  • y: is the target, which can be any AbstractVector whose element scitype is <:Continuous; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): returns a vector of Gaussian distributions given features Xnew having the same scitype as X above.

Predictions are probabilistic.

Specific metrics can also be predicted using:

  • predict_mean(mach, Xnew)
  • predict_mode(mach, Xnew)
  • predict_median(mach, Xnew)

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
+preds = predict_median(mach, X)
source

EvoTreeGaussian

EvoTreeGaussian is to be deprecated. Please use EvoTreeMLE with loss = :gaussian_mle.

EvoTrees.EvoTreeGaussian — Type

EvoTreeGaussian(;kwargs...)

A model type for constructing a EvoTreeGaussian, based on EvoTrees.jl, and implementing both an internal API the MLJ model interface. EvoTreeGaussian is used to perform Gaussian probabilistic regression, fitting μ and σ parameters to maximize likelihood.

Hyper-parameters

  • nrounds=10: Number of rounds. It corresponds to the number of trees that will be sequentially stacked. Must be >= 1.
  • eta=0.1: Learning rate. Each tree raw predictions are scaled by eta prior to be added to the stack of predictions. Must be > 0. A lower eta results in slower learning, requiring a higher nrounds but typically improves model performance.
  • lambda::T=0.0: L2 regularization term on weights. Must be >= 0. Higher lambda can result in a more robust model.
  • gamma::T=0.0: Minimum gain imprvement needed to perform a node split. Higher gamma can result in a more robust model. Must be >= 0.
  • max_depth=5: Maximum depth of a tree. Must be >= 1. A tree of depth 1 is made of a single prediction leaf. A complete tree of depth N contains 2^(N - 1) terminal leaves and 2^(N - 1) - 1 split nodes. Compute cost is proportional to 2^max_depth. Typical optimal values are in the 3 to 9 range.
  • min_weight=8.0: Minimum weight needed in a node to perform a split. Matches the number of observations by default or the sum of weights as provided by the weights vector. Must be > 0.
  • rowsample=1.0: Proportion of rows that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • colsample=1.0: Proportion of columns / features that are sampled at each iteration to build the tree. Should be in ]0, 1].
  • nbins=32: Number of bins into which each feature is quantized. Buckets are defined based on quantiles, hence resulting in equal weight bins. Should be between 2 and 255.
  • monotone_constraints=Dict{Int, Int}(): Specify monotonic constraints using a dict where the key is the feature index and the value the applicable constraint (-1=decreasing, 0=none, 1=increasing). !Experimental feature: note that for Gaussian regression, constraints may not be enforce systematically.
  • tree_type="binary" Tree structure to be used. One of:
    • binary: Each node of a tree is grown independently. Tree are built depthwise until max depth is reach or if min weight or gain (see gamma) stops further node splits.
    • oblivious: A common splitting condition is imposed to all nodes of a given depth.
  • rng=123: Either an integer used as a seed to the random number generator or an actual random number generator (::Random.AbstractRNG).

Internal API

Do config = EvoTreeGaussian() to construct an instance with default hyper-parameters. Provide keyword arguments to override hyper-parameter defaults, as in EvoTreeGaussian(max_depth=...).

Training model

A model is built using fit_evotree:

model = fit_evotree(config; x_train, y_train, kwargs...)

Inference

Predictions are obtained using predict which returns a Matrix of size [nobs, 2] where the second dimensions refer to μ and σ respectively:

EvoTrees.predict(model, X)

Alternatively, models act as a functor, returning predictions when called as a function with features as argument:

model(X)

MLJ

From MLJ, the type can be imported using:

EvoTreeGaussian = @load EvoTreeGaussian pkg=EvoTrees

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

Training data

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

mach = machine(model, X, y)

where

  • X: any table of input features (eg, a DataFrame) whose columns each have one of the following element scitypes: Continuous, Count, or <:OrderedFactor; check column scitypes with schema(X)

  • y: is the target, which can be any AbstractVector whose element scitype is <:Continuous; check the scitype with scitype(y)

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

Operations

  • predict(mach, Xnew): returns a vector of Gaussian distributions given features Xnew having the same scitype as X above.

Predictions are probabilistic.

Specific metrics can also be predicted using:

  • predict_mean(mach, Xnew)
  • predict_mode(mach, Xnew)
  • predict_median(mach, Xnew)

Fitted parameters

The fields of fitted_params(mach) are:

  • :fitresult: The GBTree object returned by EvoTrees.jl fitting algorithm.

Report

The fields of report(mach) are:

  • :features: The names of the features encountered in training.

Examples

# Internal API
 using EvoTrees
 params = EvoTreeGaussian(max_depth=5, nbins=32, nrounds=100)
 nobs, nfeats = 1_000, 5
@@ -70,4 +70,4 @@
 preds = predict(mach, X)
 preds = predict_mean(mach, X)
 preds = predict_mode(mach, X)
-preds = predict_median(mach, X)
source
+preds = predict_median(mach, X)source diff --git a/dev/search_index.js b/dev/search_index.js index c3bc04cf..291be992 100644 --- a/dev/search_index.js +++ b/dev/search_index.js @@ -1,3 +1,3 @@ var documenterSearchIndex = {"docs": -[{"location":"tutorials/ranking-LTRC/#Ranking-with-Yahoo!-Learning-to-Rank-Challenge.","page":"Ranking - Yahoo! LTRC","title":"Ranking with Yahoo! Learning to Rank Challenge.","text":"","category":"section"},{"location":"tutorials/ranking-LTRC/","page":"Ranking - Yahoo! LTRC","title":"Ranking - Yahoo! LTRC","text":"In this tutorial, we present how a ranking task can be tackled using regular regression techniques without compromising performance compared to specialized ranking learners. The data used is from the C14 - Yahoo! Learning to Rank Challenge, which can be obtained following a request to https://webscope.sandbox.yahoo.com.","category":"page"},{"location":"tutorials/ranking-LTRC/#Getting-started","page":"Ranking - Yahoo! LTRC","title":"Getting started","text":"","category":"section"},{"location":"tutorials/ranking-LTRC/","page":"Ranking - Yahoo! LTRC","title":"Ranking - Yahoo! LTRC","text":"To begin, we load the required packages:","category":"page"},{"location":"tutorials/ranking-LTRC/","page":"Ranking - Yahoo! LTRC","title":"Ranking - Yahoo! LTRC","text":"using EvoTrees\nusing DataFrames\nusing Statistics: mean\nusing CategoricalArrays\nusing Random","category":"page"},{"location":"tutorials/ranking-LTRC/#Load-LIBSVM-format-data","page":"Ranking - Yahoo! LTRC","title":"Load LIBSVM format data","text":"","category":"section"},{"location":"tutorials/ranking-LTRC/","page":"Ranking - Yahoo! LTRC","title":"Ranking - Yahoo! LTRC","text":"Some datasets come in the so called LIBSVM format, which stores data using a sparse representation: ","category":"page"},{"location":"tutorials/ranking-LTRC/","page":"Ranking - Yahoo! LTRC","title":"Ranking - Yahoo! LTRC","text":"