(Weka23PART源代码分析1.docVIP

Weka[23] PART 源代码分析 作者：Koala++/屈伟 rose 璐问我这个算法，我才去看它的论文和算法的，因为个人时间有限，分析的有些 粗糙。 请先把论文 Generating Accurate Rule Sets Without Global Optimization 看一下。 PART 在 classifiers.rules 包下面，我们直接从 buildClassifier 开始。 public void buildClassifier(Instances instances) throws Exception { // can classifier handle the data? getCapabilities().testWithFail(instances); // remove instances with missing class instances = new Instances(instances); instances.deleteWithMissingClass(); ModelSelection modSelection; if (m_binarySplits) modSelection = new BinC45ModelSelection(m_minNumObj, instances); else modSelection = new C45ModelSelection(m_minNumObj, instances); if (m_unpruned) m_root = new MakeDecList(modSelection, m_minNumObj); else if (m_reducedErrorPruning) m_root = new MakeDecList(modSelection, m_numFolds, m_minNumObj, m_Seed); else m_root = new MakeDecList(modSelection, m_CF, m_minNumObj); m_root.buildClassifier(instances); if (m_binarySplits) { ((BinC45ModelSelection) modSelection).cleanup(); } else { ((C45ModelSelection) modSelection).cleanup(); } } 如果以前看过 J48 的代码，相信看到这段代码不会陌生，不同的地方是以前是 C45PruneableClassifierTree，现在是 MakeDecList。 因为讲树分类器的次数太多(ID3，J48，NBTree，REPTree)，所以就不想再讲的太细了， 我们直接看 m_root.buildClassifier(instances)这句话。 theRules = new Vector(); if ((reducedErrorPruning) !(unpruned)) { Random random = new Random(m_seed); data.randomize(random); data.stratify(numSetS); oldGrowData = data.trainCV(numSetS, numSetS - 1, random); oldPruneData = data.testCV(numSetS, numSetS - 1); } else { oldGrowData = data; oldPruneData = null; } 如果要剪枝，就用 trainCV 和 testCV 把数据集分成 oldGrowData 和 oldPruneData，如 果不需要剪枝，那么 oldGrowData 就等于 Data，这已经在 REPTree 中讲过了。 while (Utils.gr(oldGrowData.numInstances(), 0)) { // Create rule if (unpruned) { currentRule = new ClassifierDecList(toSelectModeL, minNumObj); ((ClassifierDecList) currentRule).buildRule(oldGrowData); } else if (reducedErrorPruning) { currentRule = new PruneableDecList(toSelectModeL, minNumObj); ((PruneableDecList)