|
| 1 | +# FeatureSelection |
| 2 | + |
| 3 | +FeatureSelction is a julia package containing implementations of feature selection algorithms for use with the machine learning toolbox |
| 4 | +[MLJ](https://juliaai.github.io/MLJ.jl/dev/). |
| 5 | + |
| 6 | +# Installation |
| 7 | +On a running instance of Julia with at least version 1.6 run |
| 8 | +```julia |
| 9 | +import Pkg; |
| 10 | +Pkg.add("FeatureSelection") |
| 11 | +``` |
| 12 | + |
| 13 | +# Example Usage |
| 14 | +Lets build a supervised recursive feature eliminator with `RandomForestRegressor` |
| 15 | +from [DecisionTree.jl](https://github.com/JuliaAI/DecisionTree.jl) as our base model. |
| 16 | +But first we need a dataset to train on. We shall create a synthetic dataset popularly |
| 17 | +known in the R community as the friedman dataset#1. Notice how the target vector for this |
| 18 | +dataset depends on only the first five columns of feature table. So we expect that our |
| 19 | +recursive feature elimination should return the first columns as important features. |
| 20 | +```@meta |
| 21 | +DocTestSetup = quote |
| 22 | + using MLJ, FeatureSelection, StableRNGs |
| 23 | + rng = StableRNG(10) |
| 24 | + A = rand(rng, 50, 10) |
| 25 | + X = MLJ.table(A) # features |
| 26 | + y = @views( |
| 27 | + 10 .* sin.( |
| 28 | + pi .* A[:, 1] .* A[:, 2] |
| 29 | + ) .+ 20 .* (A[:, 3] .- 0.5).^ 2 .+ 10 .* A[:, 4] .+ 5 * A[:, 5] |
| 30 | + ) # target |
| 31 | + RandomForestRegressor = @load RandomForestRegressor pkg=DecisionTree |
| 32 | + forest = RandomForestRegressor(rng=rng) |
| 33 | + rfe = RecursiveFeatureElimination( |
| 34 | + model = forest, n_features=5, step=1 |
| 35 | + ) # see doctring for description of defaults |
| 36 | + mach = machine(rfe, X, y) |
| 37 | + fit!(mach) |
| 38 | +
|
| 39 | + rfe = RecursiveFeatureElimination(model = forest) |
| 40 | + tuning_rfe_model = TunedModel( |
| 41 | + model = rfe, |
| 42 | + measure = rms, |
| 43 | + tuning = Grid(rng=rng), |
| 44 | + resampling = StratifiedCV(nfolds = 5), |
| 45 | + range = range( |
| 46 | + rfe, :n_features, values = 1:10 |
| 47 | + ) |
| 48 | + ) |
| 49 | + self_tuning_rfe_mach = machine(tuning_rfe_model, X, y) |
| 50 | + fit!(self_tuning_rfe_mach) |
| 51 | +end |
| 52 | +``` |
| 53 | +```@example example1 |
| 54 | +using MLJ, FeatureSelection, StableRNGs |
| 55 | +rng = StableRNG(10) |
| 56 | +A = rand(rng, 50, 10) |
| 57 | +X = MLJ.table(A) # features |
| 58 | +y = @views( |
| 59 | + 10 .* sin.( |
| 60 | + pi .* A[:, 1] .* A[:, 2] |
| 61 | + ) .+ 20 .* (A[:, 3] .- 0.5).^ 2 .+ 10 .* A[:, 4] .+ 5 * A[:, 5] |
| 62 | +) # target |
| 63 | +``` |
| 64 | +Now we that we have our data we can create our recursive feature elimination model and |
| 65 | +train it on our dataset |
| 66 | +```@example example1 |
| 67 | +RandomForestRegressor = @load RandomForestRegressor pkg=DecisionTree |
| 68 | +forest = RandomForestRegressor(rng=rng) |
| 69 | +rfe = RecursiveFeatureElimination( |
| 70 | + model = forest, n_features=5, step=1 |
| 71 | +) # see doctring for description of defaults |
| 72 | +mach = machine(rfe, X, y) |
| 73 | +fit!(mach) |
| 74 | +``` |
| 75 | +We can inspect the feature importances in two ways: |
| 76 | +```jldoctest |
| 77 | +julia> report(mach).ranking |
| 78 | +10-element Vector{Int64}: |
| 79 | + 1 |
| 80 | + 1 |
| 81 | + 1 |
| 82 | + 1 |
| 83 | + 1 |
| 84 | + 2 |
| 85 | + 3 |
| 86 | + 4 |
| 87 | + 5 |
| 88 | + 6 |
| 89 | +
|
| 90 | +julia> feature_importances(mach) |
| 91 | +10-element Vector{Pair{Symbol, Int64}}: |
| 92 | + :x1 => 6 |
| 93 | + :x2 => 5 |
| 94 | + :x3 => 4 |
| 95 | + :x4 => 3 |
| 96 | + :x5 => 2 |
| 97 | + :x6 => 1 |
| 98 | + :x7 => 1 |
| 99 | + :x8 => 1 |
| 100 | + :x9 => 1 |
| 101 | + :x10 => 1 |
| 102 | +``` |
| 103 | +Note that a variable with lower rank has more significance than a variable with higher rank while a variable with higher feature importance is better than a variable with lower feature importance. |
| 104 | + |
| 105 | +We can view the important features used by our model by inspecting the `fitted_params` |
| 106 | +object. |
| 107 | +```jldoctest |
| 108 | +julia> p = fitted_params(mach) |
| 109 | +(features_left = [:x1, :x2, :x3, :x4, :x5], |
| 110 | + model_fitresult = (forest = Ensemble of Decision Trees |
| 111 | +Trees: 100 |
| 112 | +Avg Leaves: 25.26 |
| 113 | +Avg Depth: 8.36,),) |
| 114 | +
|
| 115 | +julia> p.features_left |
| 116 | +5-element Vector{Symbol}: |
| 117 | + :x1 |
| 118 | + :x2 |
| 119 | + :x3 |
| 120 | + :x4 |
| 121 | + :x5 |
| 122 | +``` |
| 123 | +We can also call the `predict` method on the fitted machine, to predict using a |
| 124 | +random forest regressor trained using only the important features, or call the `transform` |
| 125 | +method, to select just those features from some new table including all the original |
| 126 | +features. For more info, type `?RecursiveFeatureElimination` on a Julia REPL. |
| 127 | + |
| 128 | +Okay, let's say that we didn't know that our synthetic dataset depends on only five |
| 129 | +columns from our feature table. We could apply cross fold validation |
| 130 | +`StratifiedCV(nfolds=5)` with our recursive feature elimination model to select the |
| 131 | +optimal value of `n_features` for our model. In this case we will use a simple Grid |
| 132 | +search with root mean square as the measure. |
| 133 | +```@example example1 |
| 134 | +rfe = RecursiveFeatureElimination(model = forest) |
| 135 | +tuning_rfe_model = TunedModel( |
| 136 | + model = rfe, |
| 137 | + measure = rms, |
| 138 | + tuning = Grid(rng=rng), |
| 139 | + resampling = StratifiedCV(nfolds = 5), |
| 140 | + range = range( |
| 141 | + rfe, :n_features, values = 1:10 |
| 142 | + ) |
| 143 | +) |
| 144 | +self_tuning_rfe_mach = machine(tuning_rfe_model, X, y) |
| 145 | +fit!(self_tuning_rfe_mach) |
| 146 | +``` |
| 147 | +As before we can inspect the important features by inspecting the object returned by |
| 148 | +`fitted_params` or `feature_importances` as shown below. |
| 149 | +```jldoctest |
| 150 | +julia> fitted_params(self_tuning_rfe_mach).best_fitted_params.features_left |
| 151 | +5-element Vector{Symbol}: |
| 152 | + :x1 |
| 153 | + :x2 |
| 154 | + :x3 |
| 155 | + :x4 |
| 156 | + :x5 |
| 157 | +
|
| 158 | +julia> feature_importances(self_tuning_rfe_mach) |
| 159 | +10-element Vector{Pair{Symbol, Int64}}: |
| 160 | + :x1 => 6 |
| 161 | + :x2 => 5 |
| 162 | + :x3 => 4 |
| 163 | + :x4 => 3 |
| 164 | + :x5 => 2 |
| 165 | + :x6 => 1 |
| 166 | + :x7 => 1 |
| 167 | + :x8 => 1 |
| 168 | + :x9 => 1 |
| 169 | + :x10 => 1 |
| 170 | +``` |
| 171 | +and call `predict` on the tuned model machine as shown below |
| 172 | +```@example example1 |
| 173 | +Xnew = MLJ.table(rand(rng, 50, 10)) # create test data |
| 174 | +predict(self_tuning_rfe_mach, Xnew) |
| 175 | +``` |
| 176 | +In this case, prediction is done using the best recursive feature elimination model gotten |
| 177 | +from the tuning process above. |
| 178 | + |
| 179 | +For resampling methods different from cross-validation, and for other |
| 180 | + `TunedModel` options, such as parallelization, see the |
| 181 | + [Tuning Models](https://juliaai.github.io/MLJ.jl/dev/tuning_models/) section of the MLJ manual. |
| 182 | +[MLJ Documentation](https://juliaai.github.io/MLJ.jl/dev/) |
| 183 | +```@meta |
| 184 | +DocTestSetup = nothing |
| 185 | +``` |
0 commit comments