Merge pull request #1010 from mrapp-ke/use-complete-heads-by-default

Use complete heads by default

CHANGELOG.md

@@ -27,6 +27,7 @@ This release comes with several API changes. For an updated overview of the avai
 
 - The Python module or source file providing an integration with the machine learning algorithm to be used by the command line API must now be specified as described [here](https://mlrl-boomer.readthedocs.io/en/0.11.0/user_guide/testbed/arguments.html#basic-usage).
 - Several parameters and their values have been renamed to better reflect the scope of the project, which now includes multi-output regression problems. For an up-to-date list of parameters, please refer to the [documentation](https://mlrl-boomer.readthedocs.io/en/0.11.0/).
+- Rules with complete heads are now learned by default when using a decomposable loss function and a dense format for storing statistics.
 
 ## Version 0.10.2 (Aug. 9th, 2024)
 

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();
     }
 
 }

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>

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>