Merge pull request #2 from tiemvanderdeure/explain

add shapley values

Project.toml

@@ -14,6 +14,7 @@ MLJDecisionTreeInterface = "c6f25543-311c-4c74-83dc-3ea6d1015661"
 MLJGLMInterface = "caf8df21-4939-456d-ac9c-5fefbfb04c0c"
 PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
 Rasters = "a3a2b9e3-a471-40c9-b274-f788e487c689"
+Shapley = "855ca7ad-a6ef-4de2-9ca8-726fe2a39065"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 Tables = "bd369af6-aec1-5ad0-b16a-f7cc5008161c"
 

src/SpeciesDistributionModels.jl

@@ -2,17 +2,18 @@ module SpeciesDistributionModels
 
 using MLJ 
 
-import GLM, Tables, StatsBase, PrettyTables, Rasters, EvoTrees, DecisionTree
+import GLM, Tables, StatsBase, PrettyTables, Rasters, EvoTrees, DecisionTree, Shapley
 
 using Rasters: Raster, RasterStack
 
 import CategoricalArrays.CategoricalArray
 
-export SDMensemble, predict, sdm, select, machines, machine_keys
+export SDMensemble, predict, sdm, select, machines, machine_keys, shap
 
 include("models.jl")
 include("ensemble.jl")
 include("predict.jl")
+include("explain.jl")
 
 end
 

src/ensemble.jl

@@ -53,7 +53,7 @@ function select(
 end
 
 function Base.show(io::IO, mime::MIME"text/plain", ensemble::SDMensemble)
-    println(io, "SDMensemble with $(length(ensemble)) models")
+    println(io, "SDMensemble with $(Base.length(ensemble)) models")
 
     println(io, "Model performance:")
 
@@ -102,8 +102,8 @@ function sdm(
 
     @assert Tables.istable(presences) && Tables.istable(absence)
 
-    n_presence = length(Tables.rows(presences)) ##
-    n_absence = length(Tables.rows(absence))
+    n_presence = Base.length(Tables.rows(presences)) ##
+    n_absence = Base.length(Tables.rows(absence))
     n_total = n_presence + n_absence
 
     # merge presence and absence data into one namedtuple of vectors

src/explain.jl

@@ -0,0 +1,35 @@
+struct SDMshapley
+    values::Vector{<:NamedTuple} # Contains all shap values for all models
+    importances::Vector{<:NamedTuple} # Contais mean absolute shap for each variable for each model
+    ensemble::SDMensemble
+    summary
+end
+
+Base.size(shap::SDMshapley) = Base.size(shap.values)
+Base.length(shap::SDMshapley) = Base.length(shap.values)
+
+function Base.show(io::IO, mime::MIME"text/plain", shap::SDMshapley)
+    println(io, "Shapley evaluation for SDM ensemble with $(Base.length(shap)) models")
+
+    println(io, "Mean feature importance:")
+    Base.show(io, mime, shap.summary)
+end
+
+function shap(ensemble; parallelism = Shapley.CPUThreads(), n_samples = 50)
+    shapvalues = map(ensemble.trained_models) do model
+        Shapley.shapley(
+            x -> Float64.(MLJ.pdf.(MLJ.predict(model.machine, x), true)), # some ml models return float32s - where to handle this?
+            Shapley.MonteCarlo(parallelism, n_samples), 
+            ensemble.data.predictor
+        )
+    end
+
+    importances = map(vals -> map(val -> mapreduce(abs, +, val) / Base.length(val), vals), shapvalues)
+
+    summary = NamedTuple(var => mapreduce(x -> getfield(x, var), +, importances) / Base.length(importances) for var in ensemble.predictors)
+
+    return SDMshapley(shapvalues, importances, ensemble, summary)
+end
+
+
+