Fix epsilont early stopping in BoostAODE

.vscode/launch.json

@@ -106,12 +106,12 @@
             "type": "lldb",
             "request": "launch",
             "name": "test",
-            "program": "${workspaceFolder}/build_debug/tests/unit_tests",
+            "program": "${workspaceFolder}/build_debug/tests/unit_tests_bayesnet",
             "args": [
-                "-c=\"Metrics Test\"",
+                //"-c=\"Metrics Test\"",
                 // "-s",
             ],
-            "cwd": "${workspaceFolder}/build/tests",
+            "cwd": "${workspaceFolder}/build_debug/tests",
         },
         {
             "name": "Build & debug active file",

CHANGELOG.md

@@ -5,6 +5,15 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+## [1.0.2] - 2024-02-20
+
+### Fixed
+
+- Fix bug in BoostAODE: do not include the model if epsilon sub t is greater than 0.5
+- Fix bug in BoostAODE: compare accuracy with previous accuracy instead of the first of the ensemble if convergence true
+
 ## [1.0.1] - 2024-02-12
 
 ### Added

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.20)
 
 project(BayesNet
-  VERSION 1.0.1
+  VERSION 1.0.2
   DESCRIPTION "Bayesian Network and basic classifiers Library."
   HOMEPAGE_URL "https://github.com/rmontanana/bayesnet"
   LANGUAGES CXX

src/BayesNet/BoostAODE.cc

@@ -121,6 +121,9 @@ namespace bayesnet {
     }
     void BoostAODE::trainModel(const torch::Tensor& weights)
     {
+        fitted = true;
+        // Algorithm based on the adaboost algorithm for classification
+        // as explained in Ensemble methods (Zhi-Hua Zhou, 2012)
         std::unordered_set<int> featuresUsed;
         if (selectFeatures) {
             featuresUsed = initializeModels();
@@ -132,9 +135,8 @@ namespace bayesnet {
         // Variables to control the accuracy finish condition
         double priorAccuracy = 0.0;
         double delta = 1.0;
-        double threshold = 1e-4;
-        int count = 0; // number of times the accuracy is lower than the threshold
-        fitted = true; // to enable predict
+        double convergence_threshold = 1e-4;
+        int count = 0; // number of times the accuracy is lower than the convergence_threshold
         // Step 0: Set the finish condition
         // if not repeatSparent a finish condition is run out of features
         // n_models == maxModels
@@ -160,7 +162,6 @@ namespace bayesnet {
                     continue;
                 }
             }
-            featuresUsed.insert(feature);
             model = std::make_unique<SPODE>(feature);
             model->fit(dataset, features, className, states, weights_);
             auto ypred = model->predict(X_train);
@@ -169,6 +170,12 @@ namespace bayesnet {
             auto mask_right = ypred == y_train;
             auto masked_weights = weights_ * mask_wrong.to(weights_.dtype());
             double epsilon_t = masked_weights.sum().item<double>();
+            if (epsilon_t > 0.5) {
+                // Inverse the weights policy (plot ln(wt))
+                // "In each round of AdaBoost, there is a sanity check to ensure that the current base 
+                // learner is better than random guess" (Zhi-Hua Zhou, 2012)
+                break;
+            }
             double wt = (1 - epsilon_t) / epsilon_t;
             double alpha_t = epsilon_t == 0 ? 1 : 0.5 * log(wt);
             // Step 3.2: Update weights for next classifier
@@ -180,6 +187,7 @@ namespace bayesnet {
             double totalWeights = torch::sum(weights_).item<double>();
             weights_ = weights_ / totalWeights;
             // Step 3.4: Store classifier and its accuracy to weigh its future vote
+            featuresUsed.insert(feature);
             models.push_back(std::move(model));
             significanceModels.push_back(alpha_t);
             n_models++;
@@ -191,11 +199,12 @@ namespace bayesnet {
                 } else {
                     delta = accuracy - priorAccuracy;
                 }
-                if (delta < threshold) {
+                if (delta < convergence_threshold) {
                     count++;
                 }
+                priorAccuracy = accuracy;
             }
-            exitCondition = n_models >= maxModels && repeatSparent || epsilon_t > 0.5 || count > tolerance;
+            exitCondition = n_models >= maxModels && repeatSparent || count > tolerance;
         }
         if (featuresUsed.size() != features.size()) {
             notes.push_back("Used features in train: " + std::to_string(featuresUsed.size()) + " of " + std::to_string(features.size()));

tests/TestBayesModels.cc

@@ -19,7 +19,7 @@
 TEST_CASE("Library check version", "[BayesNet]")
 {
     auto clf = bayesnet::KDB(2);
-    REQUIRE(clf.getVersion() == "1.0.1");
+    REQUIRE(clf.getVersion() == "1.0.2");
 }
 TEST_CASE("Test Bayesian Classifiers score", "[BayesNet]")
 {
@@ -164,7 +164,8 @@ TEST_CASE("BoostAODE test used features in train note", "[BayesNet]")
         {"ascending",true},
         {"convergence", true},
         {"repeatSparent",true},
-        {"select_features","CFS"}
+        {"select_features","CFS"},
+        {"tolerance", 3}
         });
     clf.fit(raw.Xv, raw.yv, raw.featuresv, raw.classNamev, raw.statesv);
     REQUIRE(clf.getNumberOfNodes() == 72);