[ML] Return total SHAP per feature as a new result type (#1387)

valeriy42 · web-flow · commit 3f1b5755e54b · 2020-08-13T19:20:08.000+02:00
This PR add computation of the total feature importance values.
diff --git a/docs/CHANGELOG.asciidoc b/docs/CHANGELOG.asciidoc
@@ -75,6 +75,7 @@
   regression. (See {ml-pull}1340[#1340].)
 * Improvement in handling large inference model definitions. (See {ml-pull}1349[#1349].)
 * Add a peak_model_bytes field to model_size_stats. (See {ml-pull}1389[#1389].)
+* Calculate total feature importance as a new result type. (See {ml-pull}1387[#1387].)
 
 === Bug Fixes
 
diff --git a/include/api/CDataFrameAnalysisRunner.h b/include/api/CDataFrameAnalysisRunner.h
@@ -12,10 +12,13 @@
 
 #include <api/CDataFrameAnalysisInstrumentation.h>
 #include <api/CInferenceModelDefinition.h>
+#include <api/CInferenceModelMetadata.h>
 #include <api/ImportExport.h>
 
 #include <rapidjson/fwd.h>
 
+#include <boost/optional.hpp>
+
 #include <cstddef>
 #include <functional>
 #include <memory>
@@ -66,6 +69,7 @@ class API_EXPORT CDataFrameAnalysisRunner {
     using TProgressRecorder = std::function<void(double)>;
     using TStrVecVec = std::vector<TStrVec>;
     using TInferenceModelDefinitionUPtr = std::unique_ptr<CInferenceModelDefinition>;
+    using TOptionalInferenceModelMetadata = boost::optional<const CInferenceModelMetadata&>;
 
 public:
     //! The intention is that concrete objects of this hierarchy are constructed
@@ -141,6 +145,9 @@ class API_EXPORT CDataFrameAnalysisRunner {
     virtual TInferenceModelDefinitionUPtr
     inferenceModelDefinition(const TStrVec& fieldNames, const TStrVecVec& categoryNames) const;
 
+    //! \return A serialisable metadata of the trained model.
+    virtual TOptionalInferenceModelMetadata inferenceModelMetadata() const;
+
     //! \return Reference to the analysis instrumentation.
     virtual const CDataFrameAnalysisInstrumentation& instrumentation() const = 0;
     //! \return Reference to the analysis instrumentation.
diff --git a/include/api/CDataFrameAnalyzer.h b/include/api/CDataFrameAnalyzer.h
@@ -87,6 +87,8 @@ class API_EXPORT CDataFrameAnalyzer {
                         core::CRapidJsonConcurrentLineWriter& writer) const;
     void writeInferenceModel(const CDataFrameAnalysisRunner& analysis,
                              core::CRapidJsonConcurrentLineWriter& writer) const;
+    void writeInferenceModelMetadata(const CDataFrameAnalysisRunner& analysis,
+                                     core::CRapidJsonConcurrentLineWriter& writer) const;
 
 private:
     // This has values: -2 (unset), -1 (missing), >= 0 (control field index).
diff --git a/include/api/CDataFrameTrainBoostedTreeClassifierRunner.h b/include/api/CDataFrameTrainBoostedTreeClassifierRunner.h
@@ -10,6 +10,7 @@
 #include <core/CSmallVector.h>
 
 #include <api/CDataFrameTrainBoostedTreeRunner.h>
+#include <api/CInferenceModelMetadata.h>
 #include <api/ImportExport.h>
 
 #include <rapidjson/fwd.h>
@@ -40,6 +41,8 @@ class API_EXPORT CDataFrameTrainBoostedTreeClassifierRunner final
     static const std::string NUM_TOP_CLASSES;
     static const std::string PREDICTION_FIELD_TYPE;
     static const std::string CLASS_ASSIGNMENT_OBJECTIVE;
+    static const std::string CLASSES_FIELD_NAME;
+    static const std::string CLASS_NAME_FIELD_NAME;
     static const TStrVec CLASS_ASSIGNMENT_OBJECTIVE_VALUES;
 
 public:
@@ -70,6 +73,9 @@ class API_EXPORT CDataFrameTrainBoostedTreeClassifierRunner final
     inferenceModelDefinition(const TStrVec& fieldNames,
                              const TStrVecVec& categoryNames) const override;
 
+    //! \return A serialisable metadata of the trained regression model.
+    TOptionalInferenceModelMetadata inferenceModelMetadata() const override;
+
 private:
     static TLossFunctionUPtr loss(std::size_t numberClasses);
 
@@ -82,6 +88,7 @@ class API_EXPORT CDataFrameTrainBoostedTreeClassifierRunner final
 private:
     std::size_t m_NumTopClasses;
     EPredictionFieldType m_PredictionFieldType;
+    mutable CInferenceModelMetadata m_InferenceModelMetadata;
 };
 
 //! \brief Makes a core::CDataFrame boosted tree classification runner.
diff --git a/include/api/CDataFrameTrainBoostedTreeRegressionRunner.h b/include/api/CDataFrameTrainBoostedTreeRegressionRunner.h
@@ -10,6 +10,7 @@
 #include <maths/CBoostedTreeLoss.h>
 
 #include <api/CDataFrameTrainBoostedTreeRunner.h>
+#include <api/CInferenceModelMetadata.h>
 #include <api/ImportExport.h>
 
 #include <rapidjson/fwd.h>
@@ -51,10 +52,15 @@ class API_EXPORT CDataFrameTrainBoostedTreeRegressionRunner final
     TInferenceModelDefinitionUPtr
     inferenceModelDefinition(const TStrVec& fieldNames,
                              const TStrVecVec& categoryNameMap) const override;
+    //! \return A serialisable metadata of the trained regression model.
+    TOptionalInferenceModelMetadata inferenceModelMetadata() const override;
 
 private:
     void validate(const core::CDataFrame& frame,
                   std::size_t dependentVariableColumn) const override;
+
+private:
+    mutable CInferenceModelMetadata m_InferenceModelMetadata;
 };
 
 //! \brief Makes a core::CDataFrame boosted tree regression runner.
diff --git a/include/api/CInferenceModelMetadata.h b/include/api/CInferenceModelMetadata.h
@@ -0,0 +1,67 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the Elastic License;
+ * you may not use this file except in compliance with the Elastic License.
+ */
+#ifndef INCLUDED_ml_api_CInferenceModelMetadata_h
+#define INCLUDED_ml_api_CInferenceModelMetadata_h
+
+#include <maths/CBasicStatistics.h>
+#include <maths/CLinearAlgebraEigen.h>
+
+#include <api/CInferenceModelDefinition.h>
+#include <api/ImportExport.h>
+
+#include <string>
+
+namespace ml {
+namespace api {
+
+//! \brief Class controls the serialization of the model meta information
+//! (such as totol feature importance) into JSON format.
+class API_EXPORT CInferenceModelMetadata {
+public:
+    static const std::string JSON_CLASS_NAME_TAG;
+    static const std::string JSON_CLASSES_TAG;
+    static const std::string JSON_FEATURE_NAME_TAG;
+    static const std::string JSON_IMPORTANCE_TAG;
+    static const std::string JSON_MAX_TAG;
+    static const std::string JSON_MEAN_MAGNITUDE_TAG;
+    static const std::string JSON_MIN_TAG;
+    static const std::string JSON_MODEL_METADATA_TAG;
+    static const std::string JSON_TOTAL_FEATURE_IMPORTANCE_TAG;
+
+public:
+    using TVector = maths::CDenseVector<double>;
+    using TStrVec = std::vector<std::string>;
+    using TRapidJsonWriter = core::CRapidJsonConcurrentLineWriter;
+
+public:
+    //! Writes metadata using \p writer.
+    void write(TRapidJsonWriter& writer) const;
+    void columnNames(const TStrVec& columnNames);
+    void classValues(const TStrVec& classValues);
+    const std::string& typeString() const;
+    //! Add importances \p values to the feature with index \p i to calculate total feature importance.
+    //! Total feature importance is the mean of the magnitudes of importances for individual data points.
+    void addToFeatureImportance(std::size_t i, const TVector& values);
+
+private:
+    using TMeanVarAccumulator = maths::CBasicStatistics::SSampleMeanVar<TVector>::TAccumulator;
+    using TMinMaxAccumulator = std::vector<maths::CBasicStatistics::CMinMax<double>>;
+    using TSizeMeanVarAccumulatorUMap = std::unordered_map<std::size_t, TMeanVarAccumulator>;
+    using TSizeMinMaxAccumulatorUMap = std::unordered_map<std::size_t, TMinMaxAccumulator>;
+
+private:
+    void writeTotalFeatureImportance(TRapidJsonWriter& writer) const;
+
+private:
+    TSizeMeanVarAccumulatorUMap m_TotalShapValuesMeanVar;
+    TSizeMinMaxAccumulatorUMap m_TotalShapValuesMinMax;
+    TStrVec m_ColumnNames;
+    TStrVec m_ClassValues;
+};
+}
+}
+
+#endif //INCLUDED_ml_api_CInferenceModelMetadata_h
diff --git a/include/maths/CBasicStatistics.h b/include/maths/CBasicStatistics.h
@@ -245,7 +245,7 @@ class MATHS_EXPORT CBasicStatistics {
 
             if (ORDER > 1) {
                 T r{x - s_Moments[0]};
-                T r2{r * r};
+                T r2{las::componentwise(r) * las::componentwise(r)};
                 T dMean{mean - s_Moments[0]};
                 T dMean2{las::componentwise(dMean) * las::componentwise(dMean)};
                 T variance{s_Moments[1]};
diff --git a/include/maths/CTreeShapFeatureImportance.h b/include/maths/CTreeShapFeatureImportance.h
@@ -73,6 +73,9 @@ class MATHS_EXPORT CTreeShapFeatureImportance {
     //! Get the maximum depth of any tree in \p forest.
     static std::size_t depth(const TTreeVec& forest);
 
+    //! Get the column names.
+    const TStrVec& columnNames() const;
+
 private:
     //! Collects the elements of the path through decision tree that are updated together
     struct SPathElement {
diff --git a/lib/api/CDataFrameAnalysisRunner.cc b/lib/api/CDataFrameAnalysisRunner.cc
@@ -193,6 +193,11 @@ CDataFrameAnalysisRunner::inferenceModelDefinition(const TStrVec& /*fieldNames*/
     return TInferenceModelDefinitionUPtr();
 }
 
+CDataFrameAnalysisRunner::TOptionalInferenceModelMetadata
+CDataFrameAnalysisRunner::inferenceModelMetadata() const {
+    return TOptionalInferenceModelMetadata();
+}
+
 CDataFrameAnalysisRunnerFactory::TRunnerUPtr
 CDataFrameAnalysisRunnerFactory::make(const CDataFrameAnalysisSpecification& spec) const {
     auto result = this->makeImpl(spec);
diff --git a/lib/api/CDataFrameAnalyzer.cc b/lib/api/CDataFrameAnalyzer.cc
@@ -144,6 +144,8 @@ void CDataFrameAnalyzer::run() {
         analysisRunner->waitToFinish();
         this->writeInferenceModel(*analysisRunner, outputWriter);
         this->writeResultsOf(*analysisRunner, outputWriter);
+        // TODO reactivate once Java parsing is ready
+        // this->writeInferenceModelMetadata(*analysisRunner, outputWriter);
     }
 }
 
@@ -286,6 +288,21 @@ void CDataFrameAnalyzer::writeInferenceModel(const CDataFrameAnalysisRunner& ana
     writer.flush();
 }
 
+void CDataFrameAnalyzer::writeInferenceModelMetadata(const CDataFrameAnalysisRunner& analysis,
+                                                     core::CRapidJsonConcurrentLineWriter& writer) const {
+    // Write model meta information
+    auto modelMetadata = analysis.inferenceModelMetadata();
+    if (modelMetadata) {
+        writer.StartObject();
+        writer.Key(modelMetadata->typeString());
+        writer.StartObject();
+        modelMetadata->write(writer);
+        writer.EndObject();
+        writer.EndObject();
+    }
+    writer.flush();
+}
+
 void CDataFrameAnalyzer::writeResultsOf(const CDataFrameAnalysisRunner& analysis,
                                         core::CRapidJsonConcurrentLineWriter& writer) const {
 
diff --git a/lib/api/CDataFrameTrainBoostedTreeClassifierRunner.cc b/lib/api/CDataFrameTrainBoostedTreeClassifierRunner.cc
@@ -15,6 +15,7 @@
 #include <maths/CBoostedTreeLoss.h>
 #include <maths/CDataFramePredictiveModel.h>
 #include <maths/CDataFrameUtils.h>
+#include <maths/CLinearAlgebraEigen.h>
 #include <maths/COrderings.h>
 #include <maths/CTools.h>
 #include <maths/CTreeShapFeatureImportance.h>
@@ -41,7 +42,6 @@ const std::string IS_TRAINING_FIELD_NAME{"is_training"};
 const std::string PREDICTION_PROBABILITY_FIELD_NAME{"prediction_probability"};
 const std::string PREDICTION_SCORE_FIELD_NAME{"prediction_score"};
 const std::string TOP_CLASSES_FIELD_NAME{"top_classes"};
-const std::string CLASS_NAME_FIELD_NAME{"class_name"};
 const std::string CLASS_PROBABILITY_FIELD_NAME{"class_probability"};
 const std::string CLASS_SCORE_FIELD_NAME{"class_score"};
 
@@ -162,7 +162,9 @@ void CDataFrameTrainBoostedTreeClassifierRunner::writeOneRow(
     }
 
     if (featureImportance != nullptr) {
-        int numberClasses{static_cast<int>(classValues.size())};
+        std::size_t numberClasses{classValues.size()};
+        m_InferenceModelMetadata.columnNames(featureImportance->columnNames());
+        m_InferenceModelMetadata.classValues(classValues);
         featureImportance->shap(
             row, [&](const maths::CTreeShapFeatureImportance::TSizeVec& indices,
                      const TStrVec& featureNames,
@@ -175,20 +177,47 @@ void CDataFrameTrainBoostedTreeClassifierRunner::writeOneRow(
                         writer.Key(FEATURE_NAME_FIELD_NAME);
                         writer.String(featureNames[i]);
                         if (shap[i].size() == 1) {
-                            writer.Key(IMPORTANCE_FIELD_NAME);
-                            writer.Double(shap[i](0));
+                            // output feature importance for individual classes in binary case
+                            writer.Key(CLASSES_FIELD_NAME);
+                            writer.StartArray();
+                            for (std::size_t j = 0; j < numberClasses; ++j) {
+                                double importance{(j == predictedClassId)
+                                                      ? shap[i](0)
+                                                      : -shap[i](0)};
+                                writer.StartObject();
+                                writer.Key(CLASS_NAME_FIELD_NAME);
+                                writer.String(classValues[j]);
+                                writer.Key(IMPORTANCE_FIELD_NAME);
+                                writer.Double(importance);
+                                writer.EndObject();
+                            }
+                            writer.EndArray();
                         } else {
-                            for (int j = 0; j < shap[i].size() && j < numberClasses; ++j) {
-                                writer.Key(classValues[j]);
+                            // output feature importance for individual classes in multiclass case
+                            writer.Key(CLASSES_FIELD_NAME);
+                            writer.StartArray();
+                            for (std::size_t j = 0;
+                                 j < shap[i].size() && j < numberClasses; ++j) {
+                                writer.StartObject();
+                                writer.Key(CLASS_NAME_FIELD_NAME);
+                                writer.String(classValues[j]);
+                                writer.Key(IMPORTANCE_FIELD_NAME);
                                 writer.Double(shap[i](j));
+                                writer.EndObject();
                             }
-                            writer.Key(CDataFrameTrainBoostedTreeRunner::IMPORTANCE_FIELD_NAME);
-                            writer.Double(shap[i].lpNorm<1>());
+                            writer.EndArray();
                         }
                         writer.EndObject();
                     }
                 }
                 writer.EndArray();
+
+                for (std::size_t i = 0; i < shap.size(); ++i) {
+                    if (shap[i].lpNorm<1>() != 0) {
+                        const_cast<CDataFrameTrainBoostedTreeClassifierRunner*>(this)
+                            ->m_InferenceModelMetadata.addToFeatureImportance(i, shap[i]);
+                    }
+                }
             });
     }
     writer.EndObject();
@@ -257,6 +286,11 @@ CDataFrameTrainBoostedTreeClassifierRunner::inferenceModelDefinition(
     return std::make_unique<CInferenceModelDefinition>(builder.build());
 }
 
+CDataFrameAnalysisRunner::TOptionalInferenceModelMetadata
+CDataFrameTrainBoostedTreeClassifierRunner::inferenceModelMetadata() const {
+    return m_InferenceModelMetadata;
+}
+
 // clang-format off
 // The MAX_NUMBER_CLASSES must match the value used in the Java code. See the
 // MAX_DEPENDENT_VARIABLE_CARDINALITY in the x-pack classification code.
@@ -291,5 +325,7 @@ CDataFrameTrainBoostedTreeClassifierRunnerFactory::makeImpl(
 }
 
 const std::string CDataFrameTrainBoostedTreeClassifierRunnerFactory::NAME{"classification"};
+const std::string CDataFrameTrainBoostedTreeClassifierRunner::CLASSES_FIELD_NAME{"classes"};
+const std::string CDataFrameTrainBoostedTreeClassifierRunner::CLASS_NAME_FIELD_NAME{"class_name"};
 }
 }
diff --git a/lib/api/CDataFrameTrainBoostedTreeRegressionRunner.cc b/lib/api/CDataFrameTrainBoostedTreeRegressionRunner.cc
@@ -109,10 +109,11 @@ void CDataFrameTrainBoostedTreeRegressionRunner::writeOneRow(
     writer.Bool(maths::CDataFrameUtils::isMissing(row[columnHoldingDependentVariable]) == false);
     auto featureImportance = tree.shap();
     if (featureImportance != nullptr) {
+        m_InferenceModelMetadata.columnNames(featureImportance->columnNames());
         featureImportance->shap(
-            row, [&writer](const maths::CTreeShapFeatureImportance::TSizeVec& indices,
-                           const TStrVec& featureNames,
-                           const maths::CTreeShapFeatureImportance::TVectorVec& shap) {
+            row, [&writer, this](const maths::CTreeShapFeatureImportance::TSizeVec& indices,
+                                 const TStrVec& featureNames,
+                                 const maths::CTreeShapFeatureImportance::TVectorVec& shap) {
                 writer.Key(FEATURE_IMPORTANCE_FIELD_NAME);
                 writer.StartArray();
                 for (auto i : indices) {
@@ -126,6 +127,13 @@ void CDataFrameTrainBoostedTreeRegressionRunner::writeOneRow(
                     }
                 }
                 writer.EndArray();
+
+                for (int i = 0; i < shap.size(); ++i) {
+                    if (shap[i].lpNorm<1>() != 0) {
+                        const_cast<CDataFrameTrainBoostedTreeRegressionRunner*>(this)
+                            ->m_InferenceModelMetadata.addToFeatureImportance(i, shap[i]);
+                    }
+                }
             });
     }
     writer.EndObject();
@@ -145,6 +153,11 @@ CDataFrameTrainBoostedTreeRegressionRunner::inferenceModelDefinition(
     return std::make_unique<CInferenceModelDefinition>(builder.build());
 }
 
+CDataFrameAnalysisRunner::TOptionalInferenceModelMetadata
+CDataFrameTrainBoostedTreeRegressionRunner::inferenceModelMetadata() const {
+    return TOptionalInferenceModelMetadata(m_InferenceModelMetadata);
+}
+
 // clang-format off
 const std::string CDataFrameTrainBoostedTreeRegressionRunner::STRATIFIED_CROSS_VALIDATION{"stratified_cross_validation"};
 const std::string CDataFrameTrainBoostedTreeRegressionRunner::LOSS_FUNCTION{"loss_function"};
@@ -160,7 +173,7 @@ const std::string& CDataFrameTrainBoostedTreeRegressionRunnerFactory::name() con
 
 CDataFrameTrainBoostedTreeRegressionRunnerFactory::TRunnerUPtr
 CDataFrameTrainBoostedTreeRegressionRunnerFactory::makeImpl(const CDataFrameAnalysisSpecification&) const {
-    HANDLE_FATAL(<< "Input error: classification has a non-optional parameter '"
+    HANDLE_FATAL(<< "Input error: regression has a non-optional parameter '"
                  << CDataFrameTrainBoostedTreeRunner::DEPENDENT_VARIABLE_NAME << "'.")
     return nullptr;
 }
diff --git a/lib/api/CInferenceModelMetadata.cc b/lib/api/CInferenceModelMetadata.cc
diff --git a/lib/api/Makefile.first b/lib/api/Makefile.first
diff --git a/lib/api/unittest/CDataFrameAnalyzerFeatureImportanceTest.cc b/lib/api/unittest/CDataFrameAnalyzerFeatureImportanceTest.cc
diff --git a/lib/maths/CTreeShapFeatureImportance.cc b/lib/maths/CTreeShapFeatureImportance.cc
