Use complete heads by default when using a decomposable loss function…

… and dense statistics.
mrapp-ke · Aug 9, 2024 · cab5f66 · cab5f66
1 parent d753288
commit cab5f66
Show file tree

Hide file tree

Showing 98 changed files with 19,843 additions and 19,793 deletions.
diff --git a/cpp/subprojects/boosting/src/mlrl/boosting/rule_evaluation/head_type_auto.cpp b/cpp/subprojects/boosting/src/mlrl/boosting/rule_evaluation/head_type_auto.cpp
@@ -17,15 +17,9 @@ namespace boosting {
       AutomaticHeadConfig::createClassificationStatisticsProviderFactory(
         const IFeatureMatrix& featureMatrix, const IRowWiseLabelMatrix& labelMatrix,
         const IDecomposableClassificationLossConfig& lossConfig) const {
-        if (labelMatrix.getNumOutputs() > 1) {
-            SingleOutputHeadConfig headConfig(labelBinningConfig_, multiThreadingConfig_, l1RegularizationConfig_,
-                                              l2RegularizationConfig_);
-            return headConfig.createClassificationStatisticsProviderFactory(featureMatrix, labelMatrix, lossConfig);
-        } else {
-            CompleteHeadConfig headConfig(labelBinningConfig_, multiThreadingConfig_, l1RegularizationConfig_,
-                                          l2RegularizationConfig_);
-            return headConfig.createClassificationStatisticsProviderFactory(featureMatrix, labelMatrix, lossConfig);
-        }
+        CompleteHeadConfig headConfig(labelBinningConfig_, multiThreadingConfig_, l1RegularizationConfig_,
+                                      l2RegularizationConfig_);
+        return headConfig.createClassificationStatisticsProviderFactory(featureMatrix, labelMatrix, lossConfig);
     }
 
     std::unique_ptr<IClassificationStatisticsProviderFactory>
@@ -73,11 +67,11 @@ namespace boosting {
     }
 
     bool AutomaticHeadConfig::isPartial() const {
-        return lossConfig_.get().isDecomposable();
+        return lossConfig_.get().isDecomposable() && lossConfig_.get().isSparse();
     }
 
     bool AutomaticHeadConfig::isSingleOutput() const {
-        return lossConfig_.get().isDecomposable();
+        return lossConfig_.get().isDecomposable() && lossConfig_.get().isSparse();
     }
 
 }
diff --git a/python/subprojects/testbed/tests/res/out/boomer-classifier/binary-features-dense.txt b/python/subprojects/testbed/tests/res/out/boomer-classifier/binary-features-dense.txt
@@ -12,21 +12,21 @@ DEBUG A sparse matrix is used to store the predicted labels
 INFO Successfully predicted in <duration>
 INFO Evaluation result for test data:
 
-Example-wise F1         43.07
-Example-wise Jaccard    31.28
-Example-wise Precision  61.51
-Example-wise Recall     36.12
-Hamming Accuracy        94.68
-Hamming Loss             5.32
-Macro F1                10.58
-Macro Jaccard            7.81
-Macro Precision         88.88
-Macro Recall             9.27
-Micro F1                47.2
-Micro Jaccard           30.89
-Micro Precision         63.9
-Micro Recall            37.42
-Subset 0/1 Loss         96.98
-Subset Accuracy          3.02
+Example-wise F1         57.44
+Example-wise Jaccard    46.65
+Example-wise Precision  75.41
+Example-wise Recall     53.46
+Hamming Accuracy        95.54
+Hamming Loss             4.46
+Macro F1                22.37
+Macro Jaccard           16.81
+Macro Precision         61.04
+Macro Recall            18.92
+Micro F1                58.78
+Micro Jaccard           41.63
+Micro Precision         71.2
+Micro Recall            50.05
+Subset 0/1 Loss         84.88
+Subset Accuracy         15.12
 
 INFO Successfully finished after <duration>
diff --git a/python/subprojects/testbed/tests/res/out/boomer-classifier/binary-features-sparse.txt b/python/subprojects/testbed/tests/res/out/boomer-classifier/binary-features-sparse.txt
@@ -12,21 +12,21 @@ DEBUG A sparse matrix is used to store the predicted labels
 INFO Successfully predicted in <duration>
 INFO Evaluation result for test data:
 
-Example-wise F1         43.07
-Example-wise Jaccard    31.28
-Example-wise Precision  61.51
-Example-wise Recall     36.12
-Hamming Accuracy        94.68
-Hamming Loss             5.32
-Macro F1                10.58
-Macro Jaccard            7.81
-Macro Precision         88.88
-Macro Recall             9.27
-Micro F1                47.2
-Micro Jaccard           30.89
-Micro Precision         63.9
-Micro Recall            37.42
-Subset 0/1 Loss         96.98
-Subset Accuracy          3.02
+Example-wise F1         57.44
+Example-wise Jaccard    46.65
+Example-wise Precision  75.41
+Example-wise Recall     53.46
+Hamming Accuracy        95.54
+Hamming Loss             4.46
+Macro F1                22.37
+Macro Jaccard           16.81
+Macro Precision         61.04
+Macro Recall            18.92
+Micro F1                58.78
+Micro Jaccard           41.63
+Micro Precision         71.2
+Micro Recall            50.05
+Subset 0/1 Loss         84.88
+Subset Accuracy         15.12
 
 INFO Successfully finished after <duration>