WekaApriori源代码分析.docVIP

Weka[24] Apriori 源代码分析 作者：Koala++/屈伟 曾经卖过一个 Apriori 的程序，那个程序大约有 50%的正确率(当然结果是正确的，是只 实现上很不一样)，数据挖掘课上写了一个 Apriori，一部分懒地按书上的算法，大约对了 80% (当然结果仍然是正确的)，记得邱强有一次要用 Apriori 算法时说：weka 的太慢了，还好上 次数据挖掘课实现了一下，还挺快的，注意的一点是关联规则不属于机器学习，这里我不想 再分出来一个数据挖掘的组了。 从 buildAssociations 函数开始： double[] confidences, supports; int[] indices; FastVector[] sortedRuleSet; int necSupport = 0; instances = new Instances(instances); if (m_removeMissingCols) { instances = removeMissingColumns(instances); } 看一下 removeMissingColumns，虽然它是如此的不重要： protected Instances removeMissingColumns(Instances instances) throws Exception { int numInstances = instances.numInstances(); StringBuffer deleteString = new StringBuffer(); int removeCount = 0; boolean first = true; int maxCount = 0; for (int i = 0; i instances.numAttributes(); i++) { AttributeStats as = instances.attributeStats(i); if (m_upperBoundMinSupport == 1.0 maxCount != numInstances) { // see if we can decrease this by looking for the most frequent // value int[] counts = as.nominalCounts; if (counts[Utils.maxIndex(counts)] maxCount) { maxCount = counts[Utils.maxIndex(counts)]; } } if (as.missingCount == numInstances) { if (first) { deleteString.append((i + 1)); first = false; } else { deleteString.append(, + (i + 1)); } removeCount++; } } if (m_verbose) { System.err.println(Removed : + removeCount + columns with all missing + values.); } if (m_upperBoundMinSupport == 1.0 maxCount != numInstances) { m_upperBoundMinSupport = (double) maxCount / (double) numInstances; if (m_verbose) { System.err.println(Setting upper bound min support to : + m_upperBoundMinSupport); } } if (deleteString.toString().length() 0) { Remove af = new Remove(); af.setAttributeIndices(deleteString.toString()); af.setInvertSelection(false); af.setInputFormat(instances); Instances newInst = Filter.useFilter(instances, af); return newInst; } return instances; } For 循环中的第一个 if 不重要，不要理睬。它的作用是想在后面也就是 maxCount / numInstances 那看一下 support 的下界设多少。第二个 i